cd /news/developer-tools/warden-run-a-fleet-of-coding-agents-… · home topics developer-tools article
[ARTICLE · art-75143] src=github.com ↗ pub= topic=developer-tools verified=true sentiment=↑ positive

Warden – run a fleet of coding agents without losing your mind

Warden is a free, self-hosted Go binary that orchestrates fleets of coding agents in isolated git worktrees, tracking costs and token savings without a database or telemetry. The tool supports multiple agent backends, a live TUI and web dashboard, and an MCP server for agent-driven coordination, with an autopilot mode that autonomously manages worker agents through CI and PRs.

read74 min views1 publishedJul 27, 2026
Warden – run a fleet of coding agents without losing your mind
Image: source

📖

Docs & guide:[https://srjn45.github.io/warden/]

Run a fleet of coding agents without losing your mind. warden is a single Go binary (warden

, aliased wd

) that spawns, monitors, and tears down coding-agent sessions — each in its own isolated git worktree — while tracking exactly what every agent costs and how many tokens its lifecycle features keep out of context. It drives multiple agent backends (Claude Code by default — see Agent backends), backed by a local daemon and a file-based JSON store: no database, no SaaS, no telemetry.

Orchestrate, don't babysit— spawn many agents in parallel, watch them in a live TUI cockpit or web dashboard, and talk to any one of them. Write-type agents get their own worktree by default, so parallel agents never collide on the same tree.See what it costswarden spend

prices each agent's real model usage into dollars with a budget gate (priced for the Claude backend; bring-your-own-model backends report tokens), andwarden savings

is an append-only ledger of the tokens warden keeps out of agents' context, with a without-vs-withA/B benchmark you can screenshot.Self-hosted and free— one binary, a loopback REST API, and a JSON store on disk. Run it on your laptop or a box.** By default nothing leaves your machine**— no telemetry, no SaaS. Opt-in features (webhook/Slack alerts, remote access, andwarden savings --calibrate

) send data only where you explicitly configure them.Driven by an agent itselfwarden mcp

exposes the whole fleet as MCP tools, so an orchestrator agent (e.g. a Claude session) can spawn, query, and coordinate the fleet for you.

go install github.com/srjn45/warden/cmd/warden@latest   # or grab a release binary (see Install)
warden setup --yes      # install missing deps (tmux, claude, …); `warden doctor` just checks
warden tutorial         # guided tour of the core loop: spawn → watch → commit → tear down
warden start "review the auth module for security issues"   # spawn your first agent
warden tui              # open the cockpit

go install

gives you the CLI / daemon / TUI / MCP server. For the embeddedweb dashboard, use a[release binary]or build withmake release

— see[Install].

Architecture — one binary, multiple faces

warden daemon

is the single writer to the on-disk session store, serving a loopback REST API and running a background poller. warden ls|status|start|done|attach|send|tail

are thin HTTP clients to the daemon. warden mcp

is a stdio MCP server that bridges MCP tool calls to the same REST API, enabling an orchestrator agent session (e.g. Claude) to query agents and talk to a specific running agent. A short alias wd

(a symlink to warden

) is installed alongside it.

alias agents=warden

New here?See[docs/USAGE.md]for a task-oriented guide to running agents day to day,[docs/FEATURES.md]for a complete catalog of what warden can do, and[FEATURES.md]for the at-a-glancecoverage matrix(which of CLI / MCP / skill / web / TUI can drive each feature). The sections below cover build, install, and contributor setup.

Capability highlights from recent releases (full notes on the releases page; the complete catalog lives in docs/FEATURES.md):

Autopilot— a goal-directed, long-running autonomous mode. Author a plan file (autopilot.plan.yaml

), runwarden autopilot init

to scaffold the config, thenwarden autopilot on

to start. Amanager agent (roleautopilot

) drives a fleet ofworker agents (roleworker

, one per task) in isolated worktrees, gates their PRs through CI, and lands them intoautopilot/integration

— all without human intervention, spawning aresolver(rolebrain

) on demand to unblock a stuck worker. A guardian heal loop keeps the manager alive through stalls, a daemon-internal overwatch nudges it to tend idle/waiting workers, and a cost-tier backend ladder escalates from free (Antigravity) to subscription (Claude, Codex) when rate-limited. The kill switch iswarden autopilot off

. The switch isper-repowarden autopilot on

enables only the current repository (add--repo <root>

to target another), the enabled set is persisted so repos come back up across a daemon restart, and the manager marks its planstatus: complete

in place when it finishes so a done run is never re-run. SeeAutopilot guideandAutopilot concepts.Live config hot-reload— edit~/.warden/config.yaml

and wardenapplies it with no daemon restart: the autopilot template,auto_approve

policy, token/context guard (tokens.*

),rails.*

,model_default

,default_permission_mode

, hint gates, andnotify.*

/webhook all re-apply on the next tick or spawn. A bad editkeeps the last-good config and alerts you rather than falling back to defaults; keys that genuinely need a restart (addr

,data_dir

, timers, loop cadences) are logged as changed-but-pending. SeeConfiguration.Agent roles (— attach a named, persistent--role

)persona to an agent at spawn (warden start … --role reviewer

) or switch it on a running agent (warden set-role <id> reviewer

, which relaunches to re-inject). Five built-in roles —general

(default, no persona),orchestrator

,implementer

,auto-merger

,reviewer

— each carrying a persona plus default spawn flags (e.g.reviewer

defaults--type pr-review

,auto-merger

turns on auto-approve).warden role list

shows the catalog; the TUI new-agent form has actrl+r

role picker and the web**+ New agent** modal a Role dropdown. SeeAgent roles.Resilience & ergonomics round-upwarden recover

re-registers archived-but-alive agents (tombstone-reaper safety net); the web/tui

cockpitself-heals(validated and auto-rebuilt if wedged;warden tui --rebuild-web-cockpit

forces it);warden tui

inside an existing tmux session lays out as anative tmux window instead of erroring (--tmux-native

);wd push --force-with-lease

for safe force-pushes; rate-limit auto-resume now also answers Claude'swait-menu and monthly spend cap(rate_limit.spend_retry_interval

); and all daemon stores (sessions, pipelines, schedules, snapshots, context, mailbox) run on an embedded ScrivaDB — still no database server.Isolation guardrails (v5.0, breaking)— write-type agents (code

/docs

/website

/debug-ci

/tests

) now spawn into their own worktree by default (--in-repo

opts out), backed by PreToolUse hooks that deny-redirect rawgit

/test commands to the first-classwarden commit

/push

/sync

/check

tools. SeeLifecycle commands & boundary enforcement.Interactive mode (— a terminal REPL with a real line editor (history, a livewarden repl

)/

-command menu, Tab completion, guided argument forms, colour) that drives the fleet via deterministic/

commands (no model) or natural language (a local-LLM conductor that turns operator intent into confirmed warden tool calls without spending cloud-model tokens); optionally hosts the cockpit master pane.Pipelines, end to end— DAG pipelines are now drivable from the** MCP tools**(create/start/show/list/cancel), ship four built-in--template

starters, and supportrun_if

conditional steps.Agent sub-trees in the TUI— agents spawned by another agent nest under their parent as a collapsible sub-tree (▸ / ▾

, indented per depth); deleting a parent with live children leaves a mutedterminated tombstoneheader so the children never orphan, reaped once the sub-tree finishes.Fleet at scale— full-textwarden search

  • tags, awarden history

archive,warden export

/import

, an append-onlywarden audit log

, spawnpreset

s and variabledprompt-template

s (browse both viawarden library

), and web batch operations.Observability— per-agent metrics & performance history (warden stats

), crash/anomaly detection, the context-size guard, and webhook/Slack notifications.Token-savings ledger (— a real, append-only ledger of the tokens warden's lifecycle features keep out of agents' context, with anwarden savings

)--benchmark

A/B headline (without-vs-with warden, % reduction, $ saved) you can screenshot.Cost governance (— the REAL model spend warden measures from agents' transcripts, priced per model into dollars and rolled up per-agent/repo/day (priced for the Claude backend; bring-your-own-model backends report tokens), plus awarden spend

)budget gate that softly warns before a spawn pushes daily/weekly spend over a configured$

cap (a$

column inwarden ls

and a live per-agent cost card on the web Metrics tab round it out).Native scheduler (— opt-in cron/at triggers that fire an agent or a pipeline on the daemon's own timer; no external crontab.warden schedule

)Snapshots & insightswarden snapshot

checkpoints a worktree + transcript for rollback;warden insights

mines agent history for patterns and parallelization wins.Branch tracking (— opt-in monitor of each agent's CI status and standing vswarden branches

)origin/main

, with non-blocking inbox/desktop alerts.Extensibility— awarden plugin

system (custom task types + lifecycle hooks over JSON-stdio) and an interactiveOpenAPI/Swagger UI at/api/docs

.Web— real URL routing (/cockpit

,/tui

,/pipelines

,/metrics

,/archive

,/others

,/agent/<id>

— deep-linkable, back/forward, shareable), a Cockpit home with a Fleet header, a highlighted top-bar▢ TUI launcher that opens the literalwarden tui

full-screen(the whole three-pane cockpit — same panes, shortcuts, real shells and live agent sessions — edge-to-edge over the viewport, Ctrl+Q to exit) so you can drive the fleet from a laptop exactly as you do locally, a dedicatedMetrics tab (per-agentand fleet-total CPU / memory, per-agent context, fleet size, tokens saved — two columns on desktop, single column on mobile), dark-mode theming, global keyboard shortcuts, Cockpit agent grouping, and an Archive tab plus an Others catch-all (last).Remote access— bearer-token auth (warden token …

) and a Docker/compose deployment.First-run tutorial (— a guided walkthrough of the core loop, with a one-line nudge until you've taken (or skipped) the tour.warden tutorial

)

Go 1.26+— to build the binary (only needed forgo install

or building from source)tmux— every agent session runs in a detached tmux window** git**— worktree creation and guarded cleanup** Claude Code**(claude

on PATH) — the default agent runtime launched in each sessionAider(aider

on PATH, optional) — only needed to spawn agents with--backend aider

; bring-your-own-model (works with local Ollama models, $0)OpenCode(opencode

on PATH, optional) — only needed to spawn agents with--backend opencode

; bring-your-own-model (works with local Ollama models, $0). Install:npm install -g opencode-ai

(https://opencode.ai)(GitHub CLI) — required forgh

pr-review

sessions to check out the PR branch, and forwarden done --create-pr

Ollama(optional) — only needed if you enable the local-LLM features (local_llm

) or thewarden repl

REPL; warden falls back to Claude when it's off or unreachable

Tip:once you have thewarden

binary, runto check these dependencies andwarden doctor

to install whatever is missing (Homebrew on macOS; apt/dnf/pacman on Linux; official installers for Claude Code and Ollama).warden setup

warden setup --yes

does it non-interactively.

warden is one self-contained binary. Pick whichever fits:

Grab the archive for your OS/arch from the latest release, extract warden

, and put it on your PATH

. Released binaries have the web dashboard embedded.

curl -fsSL https://github.com/srjn45/warden/releases/latest/download/warden_8.16.3_darwin_arm64.tar.gz | tar -xz
sudo mv warden /usr/local/bin/        # or any dir on your PATH
warden --version

macOS Gatekeeper:downloaded binaries are unsigned, so the first run may be blocked. Clear the quarantine flag once:xattr -d com.apple.quarantine $(which warden)

(or right-click → Open). Building from source — option 3 — avoids this.

Every release also ships native packages (all with the wd

alias included and the web dashboard embedded):

brew install --cask srjn45/tap/warden

curl -fsSLO https://github.com/srjn45/warden/releases/latest/download/warden_8.16.3_linux_amd64.deb
sudo apt install ./warden_8.16.3_linux_amd64.deb

sudo dnf install https://github.com/srjn45/warden/releases/latest/download/warden_8.16.3_linux_amd64.rpm

yay -S warden-bin

(Adjust the version/arch in the URLs; the .deb

/.rpm

pull in tmux

and git

as recommended packages.)

go install github.com/srjn45/warden/cmd/warden@latest

This installs the warden

binary (CLI + daemon + MCP server + TUI). Note: go install

does not bundle the web dashboard (the UI is built from web/

and embedded at release time, and isn't committed to the repo). The CLI, daemon API, TUI, and MCP server all work; for the embedded web GUI use a release binary (option 1) or build from source (option 3).

git clone https://github.com/srjn45/warden.git
cd warden
make build           # CLI/daemon/TUI only → bin/warden
make release         # builds the web UI first, then embeds it → full GUI

warden supports shell completion for Bash, Zsh, Fish, and PowerShell. Generate the completion script for your shell and install it to the appropriate location:

Bash:

warden completion bash | sudo tee /etc/bash_completion.d/warden

warden completion bash > ~/.bash_completion

Zsh:

warden completion zsh | sudo tee /usr/local/share/zsh/site-functions/_warden

mkdir -p ~/.zsh/completion
warden completion zsh > ~/.zsh/completion/_warden

Fish:

warden completion fish > ~/.config/fish/completions/warden.fish

PowerShell:

warden completion powershell > warden.ps1

After installing the completion script, restart your shell or source the file for completions to take effect.

The recommended setup on macOS installs warden as an auto-starting launchd daemon (and links the Claude skill + registers the MCP server). See the next section.

Install with the script — it builds the release, installs the binary to ~/.local/bin/warden

, renders and loads the launchd plist, links the Claude skill, and registers the MCP server:

./scripts/install.sh        # or: make install

The daemon then starts automatically at login and restarts on crash (KeepAlive = true

), listening on 127.0.0.1:8765

by default.

~/.local/bin

must be on yourPATH

to runwarden

from the shell — the installer warns if it isn't.

The launchd daemon is the macOS TCC responsible process for the agents it spawns and for its own directory picker, so reads of protected folders (Downloads, Documents, Desktop, the Music/media library) surface as "warden would like to access…" prompts. Granting Full Disk Access once silences them — but macOS ties the grant to the binary's code identity, and an unsigned Go binary gets a new identity on every rebuild, which brings the prompts back.

Run the one-time setup to give the binary a stable self-signed identity:

./scripts/codesign-setup.sh   # creates a self-signed code-signing cert (once)
./scripts/install.sh          # reinstall so the binary is signed

Then grant access once: System Settings → Privacy & Security → Full Disk Access → "+" and add ~/.local/bin/warden

. Because the signing identity is stable, the grant survives future rebuilds. (install.sh

/reinstall.sh

sign automatically when the cert exists; without it they warn and leave the binary unsigned.)

Redeploy after a code change (replaces make release && ./bin/warden daemon

):

./scripts/reinstall.sh             # rebuild UI + binary, redeploy, restart
./scripts/reinstall.sh --no-build  # redeploy the existing build only

Uninstall (stops and removes the service, binary, skill link, and MCP registration; preserves your session store at ~/.warden

and the logs):

./scripts/uninstall.sh                 # or: make uninstall
./scripts/uninstall.sh --keep-binary   # leave ~/.local/bin/warden in place

Logs:

  • stdout: /tmp/warden.daemon.log

  • stderr: /tmp/warden.daemon.err

Notifications:off by default. When enabled (setnotify.enabled: true

in the config file, or the deprecatednotify: true

), the daemon posts a macOS notification when an agent enterswaiting_for_input

,idle

(stuck),orphaned

, orerrored

. These appear only when the daemon runs in your GUI login session (a terminal, or a launchduser agent); a headless/system daemon logs them instead.

The same install script detects Linux and registers a systemd user service instead of a launchd plist. It builds the release, installs the binary to ~/.local/bin/warden

, writes ~/.config/systemd/user/warden.service

, enables it, and links the Claude skill and MCP server:

./scripts/install.sh        # or: make install

The daemon starts automatically at each login session and restarts on crash (Restart=always

), listening on 127.0.0.1:8765

by default.

~/.local/bin

must be on yourPATH

— the installer warns if it isn't.

Persistent daemon:install.sh

callsloginctl enable-linger

so the systemd user service keeps running after you close your terminal or SSH session (equivalent to launchdRunAtLoad

on macOS). This requires systemd 219+ and is a no-op on most modern distros.

Redeploy after a code change:

./scripts/reinstall.sh             # rebuild UI + binary, redeploy, restart
./scripts/reinstall.sh --no-build  # redeploy existing build only

Uninstall (stops and removes the service, binary, skill link, and MCP registration; preserves ~/.warden

):

./scripts/uninstall.sh
./scripts/uninstall.sh --keep-binary   # leave ~/.local/bin/warden in place

Logs go to the systemd journal (the unit sets StandardOutput=journal

/ StandardError=journal

), so they survive reboots even where /tmp

is a tmpfs (Fedora, Arch, …):

journalctl --user -u warden -f        # follow live
journalctl --user -u warden -n 100    # last 100 lines

Notifications:off by default. Setnotify.enabled: true

in the config file to enable (the deprecatednotify: true

also works). The daemon callsnotify-send

(libnotify) when it's onPATH

; install it withapt install libnotify-bin

(Debian/Ubuntu) ordnf install libnotify

(Fedora). Degrades to log-only ifnotify-send

is not found.

The repo ships a multi-stage Dockerfile and a

. The image is a lean Alpine runtime (static

docker-compose.yml

CGO_ENABLED=0

binary plus tmux

  • git

) with the web dashboard baked in. State lives in a ~/.warden

volume so it survives container restarts, and the daemon binds 0.0.0.0:8765

for remote access.

docker build -t warden:latest .

export WARDEN_TOKEN=$(warden token generate)
docker run -d --name warden \
  -p 8765:8765 \
  -e WARDEN_TOKEN \
  -v warden-data:/home/warden/.warden \
  warden:latest

Or with compose (reads WARDEN_TOKEN

from your environment):

export WARDEN_TOKEN=$(warden token generate)
docker compose up -d        # builds the image on first run

The dashboard/API is then reachable at http://<host>:8765

; the browser prompts for the token on first load. Don't expose the port directly to the public internet — front it with Tailscale or a Cloudflare Tunnel (see Remote access).

tmux is required.warden runs every agent inside a tmux session, so the image installstmux

(andgit

, for worktree-isolated agents) even though the container's primary job is to host the daemon, REST API, and dashboard. Spawning agents that actually call Claude additionally needs theclaude

CLI and its credentials inside the container — the base image deliberately omits these to stay lean; layer them on (and mount~/.claude

) if you want the container to drive live agents rather than just manage/observe sessions.

The hook script posts lifecycle events (SessionStart

, Notification

, Stop

, SubagentStop

, SessionEnd

) to the daemon so it can update agent status in real time without polling. SessionEnd

marks the session done (terminal) when claude exits.

Merge hooks/settings.snippet.json

into ~/.claude/settings.json

. The snippet uses a __WARDEN_HOOK__

placeholder — substitute the absolute path to hooks/warden-hook.sh

in your clone first:

sed "s|__WARDEN_HOOK__|$(pwd)/hooks/warden-hook.sh|g" hooks/settings.snippet.json

sed "s|__WARDEN_HOOK__|$(pwd)/hooks/warden-hook.sh|g" hooks/settings.snippet.json > ~/.claude/settings.json

The hook fails soft — it never blocks or errors the agent, even if the daemon is down or the session is unknown.

Warden supports per-agent model selection:

Short aliases:opus

,sonnet

,haiku

,fable

Full model IDs:claude-opus-4-8

,claude-sonnet-4-6

, etc.Default:claude-sonnet-4-6

(or themodel_default

config setting)

warden start "Complex task" --model opus


warden ls  # Shows MODEL column

Warden drives Claude Code by default, but the agent layer is pluggable: pick the backend per agent at spawn time with --backend

(CLI) or the backend

param (spawn_agent

MCP tool).

Supported agents — status:warden is fully tested only withClaude Code.Codex CLIandAntigravity CLIareβ beta— live-verified state, approval, and transcript fidelity, still maturing. The remaining non-claude

backends (Aider, OpenCode, Crush, Goose, Cursor) areexperimental / work-in-progress— functionality may be reduced or unverified.

Agent Status
Claude Code ✅ Stable — fully tested, reference backend
Aider 🧪 Experimental (WIP)
OpenCode 🧪 Experimental (WIP)
Codex CLI β Beta
Crush 🧪 Experimental (WIP)
Goose 🧪 Experimental (WIP)
Cursor CLI 🧪 Experimental (WIP)
Antigravity CLI β Beta

| Backend | --backend | Tier | Notes | |---|---|---|---| Claude Code (default) | claude | A | Full fidelity — digests, savings, priced spend, resume, all permission modes | Aider | aider | A | 🧪 Experimental. Bring-your-own-model (pass --model , e.g. ollama_chat/qwen2.5-coder:3b ); structured markdown transcript ⇒ real digests; no resume, no priced spend (tokens only), runs an autonomous --message task that exits when done | OpenCode | opencode | A | 🧪 Experimental. Bring-your-own-model (pass --model , e.g. ollama/qwen2.5-coder:3b ); structured JSON transcript (via opencode export ) ⇒ real digests; resumes the worktree's last session (opencode -c ); spend tokens-only (BYO model) | Codex CLI | codex | A | β Beta. BYO provider (via Codex config / -m ); structured JSONL transcript (rollout files) ⇒ real digests; resumes dir-scoped (codex resume --last ), upgraded to exact-id via discover-then-pin; live state + approval detection; context injection via AGENTS.md ; spend tokens-only. See docs/agent-backends/codex.md | Crush | crush | A | 🧪 Experimental. BYO model (config-driven TUI; headless crush run accepts -m ); structured JSON transcript (via crush session show --json ) ⇒ real digests; resumes dir-scoped (--continue ); initial prompt auto-typed into the TUI after launch via PromptSeeder ; context injection via CRUSH.md ; spend tokens-only. See docs/agent-backends/crush.md | Goose | goose | A | 🧪 Experimental. BYO provider (GOOSE_PROVIDER /GOOSE_MODEL env); structured JSON transcript (via goose session export ) ⇒ real digests; resumes name-deterministic (goose session -r --name <id> ); no model flag on session launch; context injection via .goosehints ; spend tokens-only. See docs/agent-backends/goose.md | Cursor CLI | cursor | C | 🧪 Experimental. Hosted plan (cursor-agent , billed to your Cursor subscription); rich native permission modes (plan /ask /auto-review /force ); resumes dir-scoped (--continue ); live state + approval/trust detection; context injection via AGENTS.md . No structured transcript yet (interactive store is unreadable SQLite) ⇒ no digests; spend tokens-only. See docs/agent-backends/cursor.md | Antigravity CLI | antigravity | A | β Beta. Google-hosted free tier (agy , multi-vendor model menu); structured trajectory JSONL (incl. tool calls / files changed) ⇒ real digests; resumes dir-scoped (agy -c ); live state + approval/trust detection; context injection via AGENTS.md ; spend tokens-only. See docs/agent-backends/antigravity.md |

export OLLAMA_API_BASE=http://127.0.0.1:11434
warden start "implement the add function" --backend aider --model ollama_chat/qwen2.5-coder:3b --dir .

warden start "implement the add function" --backend opencode --model ollama/qwen2.5-coder:3b --dir .

warden start "implement the add function" --backend codex --dir .

warden start "implement the add function" --backend crush --dir .

GOOSE_PROVIDER=ollama GOOSE_MODEL=qwen2.5-coder:3b \
warden start "implement the add function" --backend goose --dir .

Backends differ in capabilities; warden degrades gracefully rather than crashing when one lacks a capability (e.g. spend shows tokens-not-dollars for a bring-your-own-model backend; rotate/handoff re-spawn fresh when resume is unavailable). It also surfaces some backends' native superpowers as first-class verbs — warden review

(Codex's diff reviewer), warden models

(Antigravity/Cursor's live model menu), and warden fork

(branch a Codex session into a new managed agent — see warden fork); see

Agent-native superpowers. See the design (

docs/superpowers/specs/2026-06-27-pluggable-agent-backends-design.md

, §5) and roadmap item #52.Warden reads all settings from a single YAML file (default ~/.warden/config.yaml

). Run warden config init

to generate a fully-commented file, edit the values, then restart the daemon; warden config

prints what's currently live. The --config <path>

flag points any command at an alternate file, and --addr <host:port>

overrides the daemon address for a single command.

Setting Default Description
addr
127.0.0.1:8765
Daemon listen address. Non-loopback requires WARDEN_TOKEN (bearer-token auth — see
data_dir
~/.warden
Directory for warden state: the embedded ScrivaDB session store (sessions-db/ , with a one-time-imported read-only JSON backup in sessions/ +closed/ ), per-agent prompt files (prompts/ ), inbox, pipelines, and metrics
claude_projects_dir
~/.claude/projects
Root of Claude Code transcript directories; the poller reads agent transcripts here to generate subjects and the context gauge
model_default
claude-sonnet-4-6
Default model for new agents (a model id or alias: sonnet /opus /haiku /fable )
default_permission_mode
auto
Default permission mode for new agents (auto /default /acceptEdits /bypassPermissions /dontAsk /plan )
notify.enabled
false
Desktop notifications when an agent needs attention
approvals
true
The approvals inbox: the daemon parses recognized Claude Code tool-permission prompts and surfaces them for answering. The web AttentionQueue shows one-click option buttons, the CLI exposes warden approvals /warden approve , and the TUI shows a pinned ⏳ Approvals row — answer it in place (i , or enter on the row, then 1 -9 ; tab cycles between waiting agents) or from the web / warden approve . Unrecognized prompts always fall back to attach
tokens.guard
true
The context-size guard: the poller reads each live agent's context-window fill from its transcript, classifies it ok /warning /critical , and shows a state-colored token figure in warden ls , the TUI row, and the web tile. Master switch for the whole guard (gauge, alert, auto-compact)
tokens.warn_alert
true
Fire a desktop notification (when notify.enabled is on) once per upward crossing into the warning or critical band
tokens.auto_compact
true
When an agent is critical and idle/waiting, auto-send /compact to reclaim its context (cooldown-guarded)
tokens.force_compact
false
When an agent goes critical while still working, interrupt it (Escape), /compact once it idles, then send the resume prompt. Destructive — discards the in-flight turn — so off by default. Per-agent override via `warden force-compact <id> on off inherit`
tokens.compact_resume_prompt
(built-in)
Message sent to a force-compacted agent once compaction lands so it resumes its work
tokens.warn
200000
Warning threshold in context tokens (inclusive lower bound). If tokens.critical is not greater than this, both reset to the defaults
tokens.critical
400000
Critical threshold in context tokens (inclusive lower bound) — the auto-/compact trigger band
auto_approve
false
Auto-answer recognized permission prompts. Bare on/off, or an allow/deny rule policy (by tool / glob / regex / paths, with per-agent overrides) — see warden auto-approve
notify.webhook_enabled / notify.webhook_url
false / (empty)
POST a JSON payload to webhook_url on attention + context-size alerts (a Slack incoming-webhook URL works out of the box); runs alongside notify.enabled
collab.enabled / collab.interval / collab.hint
true / … / true
File-conflict detection across worktrees, scan interval, and the spawn-time coordination hint
rails.isolation_guard / rails.git_redirect / rails.check_redirect / rails.git_conventions
true
Boundary-enforcement hooks (see

log.level

/ log.format

info

/ text

debug

/info

/warn

/error

) and format (text

/json

); warden daemon --log-level

/--log-format

overridelocal_llm.enabled

(+ .url

/.model

/.timeout

)false

warden repl

(its /

commands work without it)local_llm.repl

false

warden repl

mode instead of a plain shellwarden config

lists every setting, including worktree.spawn_gate

/ worktree.spawn_gate_max_agents

, tokens.budget_gate

/ tokens.budget_daily_usd

/ tokens.budget_weekly_usd

, metrics

, allow_nonloopback

, pipeline.keep_done

/ pipeline.hint

, worktree.keep_done

/ worktree.auto_prune

, the auto_restart.*

and rate_limit.*

knobs, and the REPL tier knobs (local_llm.tier

/ local_llm.escalate

/ local_llm.classifier

).

Related settings are grouped into namespaced blocks (

pipeline.*

,auto_restart.*

,collab.*

,memory.*

,branch_track.*

,rate_limit.*

,http.*

,log.*

,plugins.*

, alongside the existingrails.*

/tokens.*

/notify.*

/worktree.*

/local_llm.*

). The old flat keys (collab_enabled

,log_level

,memory_inject

, …) still load asdeprecated aliases— they work but emit a one-time deprecation warning;warden config

rewrites them into the nested form.

Config namespacing:Settings are grouped into five YAML blocks —rails

,tokens

,notify

,worktree

,local_llm

. Old flat keys (e.g.token_guard

,local_llm_url

,notify

) are still accepted as deprecated aliases and migrate to the namespaced form automatically whenwarden config init

is re-run.

Legacy env vars:the oldWARDEN_*

environment variables (e.g.WARDEN_ADDR

,WARDEN_NOTIFY

,WARDEN_TOKEN_*

) are no longer read — the daemon warns once at startup if any are still set. The per-agent IPC vars warden injects into each agent (WARDEN_SESSION_ID

,WARDEN_PIPELINE_ID

,WARDEN_JOB_ID

) are not configuration and are unaffected.

The simplest way to create an agent is to give it a plain prompt and let the system figure out the rest:

warden start "review the auth module for security issues"

In the web GUI, + New agent opens a single prompt textarea — no type or repo fields.

How it works:

Runs in the caller's directory. Prompt-spawned agents runclaude --dangerously-skip-permissions '<prompt>'

(or--permission-mode acceptEdits

with--supervised

) in the directory you invokedstart

from (or the--dir

you pass) — no per-agent directory is created. Point it elsewhere with--dir

, and include any extra repo context in the prompt itself.Type is auto-assigned. Shortly after creation the daemon classifies the prompt withclaude -p

and updates the type label. It appears as "classifying…" until then. Requiresclaude

on the daemon'sPATH

; falls back toother

if unavailable.Subject is auto-generated. Each agent has a one-line subject summarizing what it is currently working on. It is seeded from the first words of the prompt at spawn, then refreshed periodically by the poller: the poller reads the agent's Claude Code transcript (looked up under theclaude_projects_dir

config setting) or, if no transcript is found, captures the tmux pane, then asksclaude -p

for an ≤8-word phrase. Refreshes are throttled and only run when the pane content has changed.Managed worktrees still available.warden start TICKET --type development --repo …

is unchanged — see the section below.

To launch an agent in a directory other than your current one, pass --dir

:

warden start "summarize recent changes" --dir /path/to/repo

When you need a managed git worktree (e.g. a development branch tied to a Jira ticket), pass --type

. The type controls whether a git worktree is created and determines how the session is set up.

Type Worktree Notes
development
yes (new branch) Creates .worktrees/<ticket> on a new branch named after the ticket
pr-review
yes (PR branch) Detached worktree; runs gh pr checkout <PR> inside it. Requires --pr or --branch . Exempt from the write-type isolation default
analysis
opt-in (--worktree )
Runs in the repo by default; pass --worktree to get a scratch branch
spike
opt-in (--worktree )
Same as analysis
code
yes (new branch) Isolated in .worktrees/<id> ; pass --in-repo to share the repo
docs
yes (new branch) Isolated in .worktrees/<id> ; pass --in-repo to share the repo
website
yes (new branch) Isolated in .worktrees/<id> ; pass --in-repo to share the repo
debug-ci
yes (new branch) Isolated in .worktrees/<id> ; pass --in-repo to share the repo
tests
yes (new branch) Isolated in .worktrees/<id> ; pass --in-repo to share the repo
other
no Catch-all; also used for unrecognized type strings

Every write-type agent (code

/docs

/website

/debug-ci

/tests

) gets its own isolated worktree by default so parallel agents don't collide on the shared tree; pass --in-repo

to opt back into the repo root. pr-review

is exempt (it already checks out the PR branch). This isolation is what makes the boundary-enforcement hooks (see Lifecycle commands & boundary enforcement) meaningful.

By default every agent runs claude --dangerously-skip-permissions

— permission prompts are suppressed and the agent runs fully autonomously; the Notification

hook still records them as events in the session doc.

Pass --supervised

to opt into a lighter permission mode (--permission-mode acceptEdits

): file edits and common filesystem commands auto-approve, but other tools (bash writes, network calls, etc.) surface the numbered permission prompt — which the approvals inbox captures and lets you answer from the web AttentionQueue (one-click buttons), the TUI (⏳ Approvals

row → i

/1

-9

), or the CLI (warden approve

) when approvals

is on. A restored agent keeps its supervised setting.

If a worktree for the ticket already exists on disk, the spawn adopts it (reattaches claude to the existing branch) instead of erroring.

warden tui   # open the cockpit
warden       # bare invocation — same thing

warden tui

(or bare warden

) opens a tmux-composited cockpit — a dedicated tmux session with three panes: an agents list (top-left), a terminal shell for running CLI commands (bottom-left), and a full-height live detail pane (right) that opens the selected agent's interactive claude

session. Browse the list freely with

/

without disturbing the detail pane; press Enter

to open an agent in it.

Agents spawned by another agent (via the spawn_agent

MCP tool) nest under their parent as a collapsible sub-tree (▸ / ▾

, indented per depth — the same affordance pipelines use), so you can see which agents an orchestrator fanned out. Deleting a parent that still has live children keeps it as a muted terminated tombstone header (terminated · N running

) — no terminal/attach pane — so its children never vanish; the daemon reaps the tombstone once the whole sub-tree finishes.

The cockpit requires tmux ≥ 3.1 (it composites real tmux panes); if tmux isn't installed it exits with an error. Launched from inside an existing tmux session, warden detects $TMUX

and lays the cockpit out as a native tmux window in your current session instead of nesting (your own keybindings, copy-mode, and resizing apply; q

closes just the cockpit window). Force the native window with warden tui --tmux-native

, or the classic own-session cockpit with env -u TMUX warden tui

— see docs/USAGE.md §7.

The list pane polls the daemon about once a second. The daemon must be running (warden daemon

) before opening the TUI.

Keys (cockpit)

Key Action
/ or j / k
Move selection (detail pane is unaffected)
/ or h / l
Collapse / expand the pipeline or agent sub-tree under the cursor
Enter
Open the selected agent (or running pipeline job) in the right detail pane — a finished agent or tombstone shows its stored detail instead of attaching
n
New agent — opens a prompt textarea; ctrl+s to submit, esc to cancel
o
Open a directory as a group (becomes the spawn target for n )
s
Send a message to the selected agent — enter to send, esc to cancel
a
Attach — full-screen the agent's (or running job's) tmux session; press to return to the dashboardCtrl-b Enter
d
Completion digest for the selected agent — scrollable overlay; d /esc to close
i
Answer pending approvals (also enter on the ⏳ Approvals row) — 1 -9 to answer, tab for next
c
Shared-context + message-traffic inspector
r
Retry a failed / needs-attention pipeline job
x
Context-sensitive: terminate the selected agent / cancel a pipeline / close an opened dir (confirm with y )
D
Delete a stopped pipeline's record (confirm with y )
?
Toggle help overlay
Alt+t
Toggle the bottom-left master pane between Claude and a shell (both stay alive)
q
Quit and tear down the cockpit

Move focus between panes with Alt+←/→/↑/↓ (no tmux prefix); toggle the bottom-left master pane between Claude and a shell with Alt+t. See docs/USAGE.md §7 for the full cockpit guide and caveats around nested tmux.

Open the live terminal cockpit. Also launched by bare warden

with no subcommand.

warden tui
warden       # equivalent

See the Terminal UI section above for the full key reference.

Spawn a new agent session. Two modes:

Prompt mode (no --type

) — pass a quoted prompt; the type is assigned automatically:

warden start "review the auth module for security issues"
warden start "investigate why the nightly build is flaky"

Managed worktree mode (--type

required) — creates a git worktree for the ticket:

warden start PROJ-350 --type development

warden start --type pr-review --pr 1234

warden start --type debug-ci

warden start --type spike --worktree

warden start PROJ-350 --type development --repo /path/to/repo --branch my-branch

warden start "refactor the auth module" --permission-mode acceptEdits

warden start "debug the API rate limit"  # uses acceptEdits mode

warden start "quick spike" --permission-mode auto  # bypass global setting

warden set-permission-mode agent-abc123 dontAsk

Flags:

--type

— task type; omit to use prompt mode (auto-typed)--repo

— repo path (default: current directory; managed worktree mode only)--branch

— new branch name (development) or checkout target (pr-review)--pr

— PR number or URL (pr-review only)--dir <path>

— directory to run a prompt-spawned agent in (default: current directory)--worktree

— opt-in worktree for analysis/spike--in-repo

— write-type opt-out: run in the shared repo instead of an isolated worktree (ignored for pr-review)--model <model>

— per-agent model (id or aliasopus

/sonnet

/haiku

/fable

); defaults to themodel_default

config setting--role <role>

— built-in agent role:general

(default, no persona) ·orchestrator

·implementer

·auto-merger

·reviewer

. Injects the role's persona as a system-prompt addendum and fills its default spawn flags (--type

/--model

/--permission-mode

/auto-approve/tags) for any you leave unset (explicit flags still win). Seewarden role list

--tags <a,b>

— attach tags (lowercased, deduped); searchable and filterable withwarden ls --tag

--preset <name>

— seed spawn defaults from a saved preset (warden preset save

); explicit flags still override--prompt-template <name> --set VAR=value

— fill a saved prompt template (warden prompt-template save

) into the spawn prompt; repeat--set

per variable. A positional prompt still wins; free-form only (no--type

)--auto-restart

— opt this agent into daemon auto-restart on error (tuned byauto_restart_*

config)--permission-mode <mode>

— control Claude's permission level (valid modes:acceptEdits

,auto

,bypassPermissions

,default

,dontAsk

,plan

); defaults to thedefault_permission_mode

config setting (default:auto

)--supervised

— legacy alias for--permission-mode acceptEdits

; risky tools prompt and the approvals inbox surfaces them (see theapprovals

setting)

List all active agent sessions with their type, status, working directory, and subject.

warden ls

DIR

shows the base name of the agent's working directory. SUBJECT

is the auto-generated one-line summary of what the agent is currently doing (empty until the first poller refresh).

Use --json

for machine-readable output (a JSON array of full session objects; an empty fleet prints []

). Useful for scripts and for Claude driving the CLI:

warden ls --json

Other flags:

--watch

/-w

— live-update the table on every agent state change over the daemon's SSE stream (Ctrl+C to exit); combine with--json

to stream one JSON snapshot per change.--tag <tag>

— filter to agents carryingeverygiven tag (AND semantics; repeatable or comma-separated). Tags are set at spawn withwarden start --tags backend,urgent

and are part of the search haystack.

Show full detail for one session: working directory, subject, worktree, branch, PR, all events.

warden status PROJ-350

Add --json

to emit the full session as a single JSON object (including the events

array):

warden status PROJ-350 --json

Register an existing Claude session into warden.

Plain shell— finds the newest Claude conversation for the directory and resumes it under a new tmux session (claude --resume

).Inside tmux— registers the current tmux session live without relaunching claude.

warden adopt                          # newest session for cwd, resume under tmux
warden adopt --session-id <uuid>      # pick a specific Claude conversation
warden adopt --dir /path/to/project   # target a different directory

Attach your terminal to the agent's tmux session interactively.

warden attach PROJ-350

The single umbrella teardown verb. By default warden stop <TICKET>

does a full teardown: terminate the tmux + claude session, clear (archive) the record, and remove the git worktree + branch (asking for confirmation first, unless --yes

). Subtractive flags keep parts around; --pr

opens a GitHub PR first while the agent is still intact. Safe order is always PR → terminate → clear record → remove worktree, so a failed push leaves the agent running.

warden stop PROJ-350                 # full teardown (asks before removing the worktree)
warden stop PROJ-350 --yes           # ...without the confirmation prompt
warden stop PROJ-350 --keep-worktree # terminate + clear record, keep the worktree (== `done`)
warden stop PROJ-350 --keep-record   # terminate + remove worktree, keep the record
warden stop PROJ-350 --hard          # purge the record instead of archiving
warden stop PROJ-350 --pr --base main # open a GitHub PR first, then tear down

The four older verbs are kept as thin aliases — each is just stop

with a fixed flag combo:

old verb equivalent
wd terminate <T>
wd stop <T> --keep-record --keep-worktree
wd delete <T> [--hard]
wd stop <T> --keep-worktree (record only)
wd remove-worktree <T>
wd stop <T> --keep-record (worktree only)
`wd done <T> [--hard --create-pr]`
`wd stop <T> --keep-worktree [--hard --pr]`
wd stop <T>
terminate + clear record + remove worktree

Terminate the agent (kill its tmux + claude session) and clear its stored record in one step. It does not remove the git worktree — that is a separate, explicitly-confirmed step (remove-worktree

). Equivalent to terminate

followed by delete

, i.e. stop --keep-worktree

.

warden done PROJ-350          # terminate + clear record (worktree kept)
warden done PROJ-350 --hard   # purge the record instead of archiving it
warden done PROJ-350 --create-pr --base main   # push the branch + open a GitHub PR first

--create-pr

pushes the agent's branch and opens a GitHub PR (via gh

) — titled from the agent, bodied from its digest, targeting --base

(default main

) — before terminating, so a failure leaves the agent running to retry; an existing PR for the branch is reported, not re-created.

Stop an agent: kill its tmux + claude session, but keep the record and worktree. This is the safe "stop this agent" default — it is reversible with warden restore

. Alias for stop --keep-record --keep-worktree

.

warden terminate PROJ-350

Recreate and resume a lost/orphaned agent's tmux + claude session (claude --resume

). Use only when the agent's tmux session is gone (status orphaned

).

warden restore PROJ-350

The safety net for the tombstone reaper (which auto-archives a parent record once its whole sub-tree goes terminal): a stale orphaned

status racing a daemon restart could previously let a genuinely-live session's record get archived out from under it. Bare warden recover

scans archived records and only reports the ones whose tmux session is confirmed still alive; --apply

re-inserts each candidate into the active store under its original id — any children (linked via parent_id

, untouched by archiving) reconnect automatically.

warden recover                # report candidates only (dry run)
warden recover --apply        # actually revive them
warden recover --json         # scripting

Clear an agent's stored record (archives by default; --hard

purges). Does not touch tmux or the worktree. Alias for stop --keep-worktree

(record only).

warden delete PROJ-350
warden delete PROJ-350 --hard

Remove an agent's git worktree and branch. Destructive. It refuses if the agent is still running (terminate it first) or if the worktree has uncommitted changes or unpushed commits — use --force

to override the guard. Alias for stop --keep-record

(worktree only); always asks unless --yes

.

warden remove-worktree PROJ-350
warden remove-worktree PROJ-350 --force

Type a message into the agent's claude session and press Enter.

warden send PROJ-350 "run the unit tests and fix any failures"

Print the recent terminal output of the agent's claude session.

warden tail PROJ-350
warden tail PROJ-350 --lines 80

Summarize what an agent accomplished — files touched, branch, number of turns, and a short narrative (best-effort, via claude -p

). Also available as a web Digest panel and, in the cockpit, the d

key (opens a scrollable digest for the selected agent).

warden digest PROJ-350
warden digest PROJ-350 --json

The approvals inbox (on by default; the approvals

setting). When a --supervised

agent hits a tool-permission prompt, the daemon recognizes it and surfaces the numbered options so you can answer without attaching.

warden approvals                 # list pending permission prompts (with their options)
warden approve PROJ-350 1  # answer prompt for that agent with option 1 (e.g. "Yes")

Unrecognized prompts always fall back to attach. Also surfaced in the web AttentionQueue (one-click buttons) and the TUI ⏳ Approvals row.

Preflight checks — required binaries (tmux

, git

, claude

), optional ones (gh

, ollama

, warn-only), daemon reachability, and the data directory. It also prints a one-line hardware-aware local_llm_model

recommendation for the REPL.

warden doctor

Verifies the install with the same checks as doctor, then installs whatever is missing — idempotent, so it only touches deps that aren't already on PATH. It prints the exact install command for each missing dependency and prompts before running it;

--yes

installs everything without prompting (for automation). Required deps (tmux

, git

, claude

) come first, then optional ones (gh

, ollama

). Package managers are auto-detected — Homebrew on macOS (never auto-bootstrapped) and apt

/dnf

/pacman

on Linux — and Claude Code and Ollama use their official installers. After installing, it re-runs the checks and prints a doctor-style report. setup

is CLI-only(it installs host packages) and is not exposed over MCP.

warden setup            # confirm-each install of anything missing
warden setup --yes      # non-interactive: install all missing deps

Recommends local models for the REPL (warden repl

), sized to this machine. It auto-detects two figures from the same memory pool — total memory (the bound) and average free memory (sampled to smooth spikes) — using NVIDIA VRAM (nvidia-smi

), Apple unified memory, or Linux MemAvailable

. It scores a curated, tool-calling-forward catalog (Qwen3, gpt-oss, Mistral Small, Qwen2.5) by conductor suitability — not raw size or coding skill, since the REPL routes tool calls and never writes code. Scores are calibrated against the Berkeley Function-Calling Leaderboard (BFCL v4), weighted toward the multi-turn subcategory that matches the REPL's tool-call loop. Each model is marked fits now

/ free memory first

/ too large

. The ★ pick is the best-scoring model that runs comfortably now with headroom for your real workload (Docker, DBs, IDE, Claude sessions, the daemon). It only ever recommends — you set local_llm_model

yourself.

warden llm suggest                    # auto-detect and rank
warden llm suggest --samples 8        # average more free-memory samples
warden llm suggest --total-gb 48 --free-gb 30   # what-if for another machine
warden llm suggest --json

A guided tour of the core loop (spawn → watch → commit → tear down). Until you've taken or skipped it, warden prints a single non-blocking stderr hint nudging you toward it (suppressed for piped/non-interactive use).

warden tutorial                       # run the walkthrough, then mark it complete
warden tutorial --skip                # mark complete without running it
warden tutorial --reset               # clear the marker so the tour (and hint) run again

Disable the hint entirely with tutorial: false

in the config.

Run inside an agent session to retire a long-lived, context-heavy agent and hand off to a fresh successor in the same workdir/worktree. Phase 1 is driven by the /warden

skill (the agent writes a handoff file + resume prompt and shows you); on your go-ahead it spawns the successor and reaps itself.

warden handoff --retire --confirm \
  --resume-file "${TMPDIR:-/tmp}/warden-rotate-handoff-$WARDEN_SESSION_ID.md" \
  --resume-prompt "Continue the migration from where the notes leave off"

The handoff file lives at a unique, per-agent temp path so concurrent rotations never clobber each other; the successor deletes it once read (and /tmp

self-clears). Spawn-before-reap is fail-safe: if the successor fails to spawn, the current agent keeps running. Rotation reuses the worktree by cwd and never removes it. --retire

is mutually exclusive with --to

.

Define a DAG of agent jobs in YAML and let the daemon run them: jobs with no dependencies start first, and each job's emit

publishes its output and unblocks its dependents. The daemon owns the cheap "await + fire" so the lead Claude stays off the critical path.

warden pipeline validate -f review.yaml # check the spec (DAG/refs/cycles); exit 0/1, no daemon
warden pipeline create -f review.yaml   # validate + register (does not start)
warden pipeline list-templates          # list the built-in starter templates + their placeholders
warden pipeline create --template analyze-implement-review --set REPO=. # render from a template
warden pipeline start <id>              # spawn jobs with no dependencies
warden pipeline show <id>               # jobs, status, branches, and emitted output
warden pipeline list
warden pipeline edit-job <id> <job> ... # edit a not-yet-started job's fields
warden pipeline retry <id> <job>        # re-run a failed/needs-attention job
warden pipeline  <id>              # stop spawning new jobs (in-flight keep running)
warden pipeline resume <id>             # resume a d pipeline
warden pipeline cancel <id>             # terminate running jobs
warden pipeline delete <id>             # remove the record (cancel first if live)

Four go:embed

-bundled templates ship in the binary — analyze-implement-review

, parallel-tasks

, test-fix-verify

, research-synthesis

. Render one with warden pipeline create --template <name>

, substituting placeholders via {{NAME}}

/{{REPO}}

(auto-filled) and --set KEY=VALUE

.

A minimal analyze → implement → review

spec (job prompts must not mention emit

— the daemon auto-appends it and auto-injects upstream outputs):

name: auth-refactor
jobs:
  - id: analyze
    prompt: "Analyze the auth module and list the concrete refactors needed."
  - id: implement
    depends_on: [analyze]
    worktree: fresh
    prompt: "Implement the refactors identified upstream."
  - id: review
    depends_on: [implement]
    prompt: "Review the implementation branch for correctness and regressions."

Jobs can be made conditional with run_if: success

(default) | failure | always

— e.g. a run_if: failure

rollback/notify step that only runs when an upstream job fails. Pipelines have full TUI and web visibility (a ▸ Pipelines section / a Pipelines tab). See docs/USAGE.md for the full authoring guide.

Fire an agent spawn or a pipeline on the daemon's own timer — no external crontab. Opt-in: set scheduler_enabled: true

in the config and keep the daemon running (schedules only fire while it is up).

warden schedule create daily-review --cron "0 9 * * *" \
  --type pr-review --repo . --prompt "Review yesterday's merged PRs"

warden schedule create launch --at 2026-06-27T09:00 --prompt "Kick off the release checklist"

warden schedule create nightly --cron "0 2 * * *" --pipeline ci.yaml

warden schedule list                  # kind, mode, spec, enabled, next run, last error
warden schedule delete daily-review

Missed runs are not backfilled — on daemon startup each next-fire is recomputed from the wall clock. The reconcile loop fails soft (a bad fire is recorded in last_error

, never crashes the loop). list_schedules

exposes the same read-only view over MCP.

The substrate pipelines are built on, usable directly so agents can collaborate:

warden ctx set build.status "green" --as agent-4f2a
warden ctx get build.status
warden ctx list --prefix build.

warden msg send agent-9c1d "the API contract changed — re-check your client"
warden msg inbox --as agent-9c1d
warden msg wait --as agent-9c1d --timeout 120   # block in the daemon until a message arrives

The daemon watches each active agent's worktree (fsnotify, with a git diff

poll as a safety net) and warns — via the inbox, deduplicated — when two agents edit the same file. Spawned agents also get a system-prompt hint to check before editing shared files, so they coordinate rather than overwrite.

warden collab conflicts                 # current cross-agent file conflicts
warden collab who-is-editing <file>     # which agents (if any) are touching a file

Tunable via the collab.enabled

/ collab.interval

/ collab.hint

config settings; also exposed as the get_collaboration_status

/ who_is_editing_file

MCP tools and a File conflicts card on the dashboard.

First-class, deterministic commands that move git and test/lint/build work off the agent and return compact results instead of raw tool-spam. PreToolUse hooks (config-gated, fail-open) steer agents toward these and deny-redirect the raw escapes — see Lifecycle commands & boundary enforcement.

warden commit            # stage + commit the agent's worktree (message auto-filled if omitted)
warden commit -m "fix: …"
warden push              # push the worktree's branch
warden push --force-with-lease  # safe force after a rebase/amend
warden sync              # rebase-sync against the upstream (refuses on a dirty tree)
warden check [name]      # run the project's .warden/check.yml checks; reports only failures

Rails: no commit/push on main

/master

, no dirty-tree sync, pre-commit-hook failures surfaced as a result. Force pushes are always --force-with-lease

(never a bare --force

), so a rebased branch can't clobber a teammate's push. All four are also MCP tools.

Some backends ship native capabilities Claude Code doesn't, and warden surfaces them as first-class verbs (added on top — never a restriction). Like warden check

, these exec in the agent's worktree with no daemon round-trip, so they are CLI-only by design (no MCP twin).

warden review                  # the backend reviews its OWN uncommitted diff and streams findings
warden review --base main      # review the branch's changes against a base instead
warden review --prompt "focus on error handling"
warden review --json           # neutral machine-readable findings: {summary, verdict, findings[]}
warden models                  # the backend's LIVE model menu (one id per line; --json for an array)

— the agent-native counterpart towarden review

warden check

(configured test/lint) and apr-review

agent (a whole reviewer session): it runs the backend's own one-shot reviewer against the worktree.Codex implements it (codex review

); backends without a native reviewer (e.g. Claude) exit non-zero pointing you atwarden check

/pr-review

.--json

runs the structured form (codex exec review

) and normalizes the backend's native output into one neutral findings shape; review quality rides the backend's configured model.— the live runtime model menu (vs warden's staticwarden models

opus

/sonnet

/haiku

/fable

aliases).Antigravity(agy models

) andCursor(cursor-agent --list-models

) implement it; the ids feed--model

verbatim. Listing is a metadata read, so it spends no quota. Backends with a static model set (Claude) degrade non-zero ("pass--model

with a known id").

Both take --backend <id>

to target a specific backend (default: the current agent's). See Agent backends.

warden owns one committed, backend-neutral project memory.warden/memory.md

(beside .warden/check.yml

), keyed implicitly by the repo root and holding durable cross-agent facts ("where X lives", "run Y via warden check

", project invariants). The point: the next agent — any backend — doesn't re-pay the rediscovery tax the last one already paid.

warden memory                  # show the resolved path + the rendered view injected into agents
warden memory --raw            # print the file verbatim
warden memory --edit           # open it in $EDITOR (auto-creates it first)
warden memory --path           # just the resolved path (scriptable; no auto-create)

At every spawn warden projects the budgeted, navigational render into the agent's system prompt through the same seam the collab/pipeline/git hints ride: Claude via --append-system-prompt

(file-backed, so it never bloats the launch line), codex / cursor / opencode / antigravity via their AGENTS.md

warden block, crush via CRUSH.md

, goose via .goosehints

; aider degrade-skips. 7 of 8 backends project with zero new adapter code. It is config-gated by memory.inject

(default on) — off, or an empty/absent file, makes the launch byte-identical to no injection. warden reads but never rewrites your CLAUDE.md/AGENTS.md/CONVENTIONS.md; .warden/memory.md

is warden's own. warden memory

is CLI-local like warden check

/warden review

(no daemon round-trip, no MCP twin).

Auto-curation ( memory.curate, default OFF). warden can also

proposememory for you: on the existing completion-digest hook, a debounced extraction pass reads finished agents' digests + the current memory and writes

durable, reusable facts back as

- [unverified · <date> · <provenance>] <fact>

. It is deliberately never authoritative — proposals land in the working tree only(warden

never commits and never pushes them, so the committed diff is the human review gate), promote to

trusted

only when corroborated by a second agent or a human, supersede older contradicting facts (struck with a tombstone),

age out past a TTL, and get

flagged stale when a named path vanishes. It prefers the

$0

local model and never sits on a paid critical path. This is the core defence against memory poisoning: one agent's wrong belief can never silently mislead the fleet.

Local grounding ( memory.ground, default ON). In

warden repl

you can also askthis memory a question —

/memory <q>

(/mem

//ask

), or the model-callable project_memory

tool — and warden answers "where does X live?" / "how do I run Y?" locally from

.warden/memory.md

. This is the token-removinglever of the feature: unlike projection (which

addsinput tokens per turn), grounding

removesa cloud round-trip by serving the answer from the local model. It is read-only (never creates or writes memory), cites each entry's trust (

unverified

/trusted

/human

) + provenance so a stale hint reads as a hint, stays on the local tier only(structurally

$0

— it can never escalate to a paid model), and with no local model configured degrades to returning the matching entries verbatim. An absent/empty file answers "not in project memory".

warden search auth backend          # AND every term across active sessions; --closed folds in the archive
warden history --since 7d --type development --limit 20   # browse the archived (closed) store

search

matches case-insensitively over id/name/ticket/type/subject/prompt/branch/tags/last-pane; --json

on either prints raw records. The web dashboard carries a live search box and a 🗄 Archive tab mirrors history

.

Serialize session metadata (not worktrees/branches/tmux) to JSON for backup or moving between machines.

warden export > fleet.json           # active records; --all also dumps the archive
warden import < fleet.json           # idempotent by id; --merge overwrites colliding records

An append-only trail of the daemon's meaningful actions (spawn

, terminate

, delete

, approve

, pipeline_start

/pipeline_cancel

) written to ~/.warden/audit.jsonl

(0600

). Read directly from the file, so it works even while the daemon is down.

warden audit log                                  # newest last (default tail 50)
warden audit log --action spawn --since 24h --json
warden audit log --target PROJ-350 --tail 0       # 0 = all

Save reusable spawn defaults under a name and replay them with warden start --preset <name>

.

warden preset save fast --model haiku --permission-mode acceptEdits --auto-restart
warden preset list
warden start "quick fix" --preset fast            # explicit flags still override the preset

--type

/--model

/--permission-mode

/--auto-restart

/--worktree

/--in-repo

are persisted to ~/.warden/presets.yaml

; per-invocation inputs (ticket, branch, PR, dir) are not.

Where a preset stores reusable flags, a prompt template stores a reusable prompt body with {{VAR}}

placeholders. Save one, then fill it in at spawn time (alias pt

).

warden prompt-template save bugfix --prompt "Fix the bug in {{FILE}} described by {{TICKET}}"
warden prompt-template list                        # each template + its variables
warden start --prompt-template bugfix --set FILE=server.go --set TICKET=WARD-42

Variables are auto-derived from the body and persisted to ~/.warden/prompt-templates.yaml

. Every declared variable must be supplied (a typo'd --set

is rejected), and an explicit positional prompt still wins. --prompt-template

is free-form only (no --type

).

One umbrella over all three kinds of reusable launch config — saved spawn presets, saved prompt templates, and the built-in pipeline templates.

warden library list                               # presets + prompt templates + pipeline templates, labeled sections (alias: wd lib list)
warden library save-preset fast --model haiku     # delegates to `warden preset save`
warden library save-prompt bugfix --prompt "Fix {{FILE}}"   # delegates to `warden prompt-template save`

Purely additive: it reuses the existing preset store, the prompt-template store, and the embedded template catalog, so warden preset

, warden prompt-template

, and warden pipeline list-templates

keep working unchanged. Pipeline templates are embedded/read-only (no save-template

; author one with warden pipeline create -f <spec.yaml>

). Also exposed over MCP as library_list

(returns {presets, prompt_templates, templates}

).

CLI view of the resource metrics (per-agent process-tree RSS/CPU, system memory/pressure, daemon self-stats).

warden stats                          # live snapshot
warden stats --history                # per-agent performance history + anomaly warnings
warden stats --history --agent PROJ-350

One umbrella over warden's two financial views: spend (the real dollars agents billed to the model provider) and savings (the tokens — and the dollars they represent — warden kept out of context). warden cost

with no subcommand prints a combined at-a-glance summary of both.

warden cost                           # combined: SPEND section + SAVINGS section
warden cost spend                     # same as `warden spend` (all flags wired through)
warden cost savings --benchmark       # same as `warden savings --benchmark`

Purely additive: it reuses the existing spend rollup and savings ledger and their render logic, so the top-level warden spend

and warden savings

keep working unchanged (they remain available as aliases). Resource footprint — memory/CPU/pressure — is a different axis; see warden stats

.

Read back the token-savings ledger — the tokens warden's lifecycle features kept out of agents' context windows. A real, append-only record, not an estimate. Gated by the savings

config setting (default on). Also reachable as warden cost savings

.

warden savings                        # per-feature table (saved/raw tokens, events)
warden savings --benchmark            # headline A/B: without-vs-with warden, % cut, $ saved, trend sparkline
warden savings --since 7d             # scope to a window (24h/7d/2w) or a date
warden savings --json                 # structured summary
warden savings --audit                # raw-vs-kept provenance samples (needs savings_samples)
warden savings --calibrate            # measure this workload's bytes/token vs Claude count_tokens (needs ANTHROPIC_API_KEY)

Two axes are reported separately and never blended: the context axis (how much leaner context stayed, in % and $) and the offload axis (cloud-model work moved off entirely onto the local LLM, in $). Each figure states its basis — CALIBRATED

or the 4-bytes/token HEURISTIC

. See docs/FEATURES.md §29.

The cost side of the ledger: the REAL billed model spend warden measured from agents' transcripts, priced per model into dollars and rolled up per agent / repo / day. Dollar pricing currently covers the Claude backend; bring-your-own-model backends report tokens only. Gated by the same savings

config setting. Also reachable as warden cost spend

.

warden spend                          # total / today / this week, then per-agent/repo/day $ tables
warden spend --by agent               # just one rollup: agent, repo, or day
warden spend --json                   # structured rollup

A budget gate (off by default) turns this into a guardrail: set budget_daily_usd

/ budget_weekly_usd

and flip budget_gate: true

, and a spawn that would push measured spend over the cap warns first (re-run with --force

to proceed) — mirroring the memory-pressure spawn gate. warden ls

also gains a COST column, and the web Metrics tab a live per-agent cost card. See docs/FEATURES.md §30.

Opt-in, read-only view of each active agent's branch health: its GitHub CI status (latest gh run list

in the worktree) and its standing vs origin/main (commits behind/ahead, merged?).

warden branches                       # table of per-agent CI + base-branch standing
warden branches --json

The daemon monitor behind it (enable with branch_track.enabled

) delivers non-blocking alerts — an inbox note to the agent (and a desktop ping to you) on a new CI failure, an inbox nudge on a merged or far-behind branch. Every gh

/git call fails open. Also exposed via GET /api/v1/collab/branches

and the get_branch_status

MCP tool.

Mine archived agent history for patterns — recurring task shapes, slow/failure-prone work, and parallelization opportunities — as a deterministic report (optionally narrated by the local LLM). Gated by insights

(default on).

warden insights
warden insights --json

Checkpoint an agent's worktree changes + session transcript and roll back later. Gated by snapshots

(default on).

warden snapshot create [name] -m "before risky refactor"   # capture a checkpoint
warden snapshot list [name] [--all]                        # list checkpoints
warden snapshot restore <id> [--force]                     # re-apply onto its worktree

Restore reapplies the captured stash onto the recorded worktree; it refuses a dirty/conflicting tree rather than clobbering, and a failed apply leaves the snapshot intact. Also available as the snapshot_create

/snapshot_list

/snapshot_restore

MCP tools.

Two layers control auto-approval:

Per-agent toggle — opt one agent into evaluation even when the global policy is off:

warden auto-approve agent-abc123 on
warden auto-approve agent-abc123 off

Rule policy — a real allow/deny engine. A recognized prompt is auto-answered only when it matches an allow rule, matches no deny rule, and isn't on warden's built-in destructive deny-list (which always wins). Rules match by tool name, a case-insensitive glob/substring (--pattern

), a Go regexp (--regex

) over the prompt, and/or path globs (--paths

). A per-agent override (--agent

, keyed by name or id) gets its own rule set that replaces the default. Changes take effect immediately and persist to config.

warden auto-approve rules                         # show the live policy
warden auto-approve enable                         # turn the policy on
warden auto-approve allow --tool Read              # auto-approve all Read prompts
warden auto-approve allow --regex '^Bash\(git (status|diff|log)\)'
warden auto-approve deny  --tool Bash --pattern rm
warden auto-approve allow --agent reviewer --tool Grep
warden auto-approve clear --agent reviewer         # drop reviewer's overrides

With no rules configured, an enabled policy keeps the simple legacy behavior: it auto-answers every recognized, non-destructive prompt (selecting the least-privilege affirmative). Multi-select / text-entry / unrecognized prompts always fall back to manual. Also available as the set_auto_approve

(toggle) and set_auto_approve_policy

(rules) MCP tools.

Change a running agent's permission mode (acceptEdits

/auto

/bypassPermissions

/default

/dontAsk

/plan

); preserved on restore.

warden set-permission-mode agent-abc123 dontAsk

A role is a named, persistent system-prompt persona attached to an agent, plus a set of default spawn flags. Every agent has exactly one role; the default is general

, which injects no persona and behaves exactly as agents do today. The role set is a fixed built-in catalog (no user-defined roles):

Role Persona Default flags
general
(none — plain agent)
orchestrator
coordinates a fleet of warden agents; plans and delegates, doesn't write feature code itself unless trivial --permission-mode auto
implementer
implements a task end-to-end on its own branch (code, tests, checks, commit, PR) --type development
auto-merger
owns getting an open PR merged: watches CI, fixes failures/conflicts, merges when green --permission-mode auto , auto-approve on
reviewer
reviews a branch/PR for correctness, coverage, and style; produces findings + a verdict, no fixes unless asked --type pr-review

The persona is injected through the same system-prompt seam warden already uses for its collab/git/pipeline hints (Claude via --append-system-prompt

, the injecting backends via their rules file); only the role name is persisted, so the persona re-resolves from the registry at every (re)launch — nothing persona-shaped is stored on disk. Role default flags fill only the fields you left unset (explicit start

flags win); default tags are unioned in.

warden role list

warden start "review PR 1234 for correctness" --role reviewer

warden set-role agent-abc123 reviewer

warden set-role agent-abc123 general

Roles are also drivable from the UIs (TUI new-agent ctrl+r

picker, web + New agent Role dropdown) and over MCP (spawn_agent

's role

param, set_role

, list_roles

) — see Orchestrator (MCP). set-role

relaunches the agent (its in-flight turn is discarded) because a persona only takes effect at (re)launch, mirroring how set-permission-mode

persists a value but a persona additionally needs a fresh launch.

The umbrella for warden's worktree operations — list and prune:

warden worktree                       # list warden-owned worktrees (same as `worktree list`)
warden worktree list                  # list worktrees under .worktrees, joined to records (alias: ls)
warden worktree prune                 # reclaim orphaned worktrees (prompts; --force overrides guards)
warden worktree prune --include-archived  # widen scope to archived records

warden prune

remains as a top-level alias for warden worktree prune

(same flags, prompts, and output). Retention is policy-driven via the worktree.keep_done

/ worktree.auto_prune

config settings; dirty/unpushed worktrees are always kept. Reclaiming whole orphaned worktrees is distinct from warden remove-worktree

, which tears down one agent's worktree.

The single verb for passing work to another agent, with three modes. Default mode spawns a fresh delegate in its own isolated worktree; --to <id>

delivers the handoff into an existing agent's inbox (waking it) — both keep the source running and inline the handoff content into the recipient's prompt/message. --retire

(requires --confirm

) is the self-succession mode: it spawns a successor in the calling agent's same worktree and reaps the caller — exactly what the warden rotate

alias runs (see above). --retire

and --to

are mutually exclusive. Phase 1 (writing the handoff) is /warden

-skill-driven.

Shorthand for warden start --fork-from <agent>

— fork an existing agent's recorded session into a new managed agent. A fork branches the source's conversation/reasoning (Codex's session rollout) into a divergent session and continues it as its own agent: a fresh sibling worktree off the source's branch HEAD, seeded with the source's uncommitted tracked changes (dirty-tree carry; untracked/.gitignore

'd build artifacts are not carried), with its own tmux session warden manages and tears down. The source agent keeps running, untouched — fork branches sideways, unlike snapshot

(rewinds one timeline) or rotate

/handoff

(carry the task, drop the conversation).

warden fork agent-7                  # fork agent-7, continue its conversation
warden fork agent-7 "now try X"      # fork and seed a divergent first prompt

Only backends with a native session fork are forkable — Codex today (codex fork

); forking one without (e.g. Claude) reports a clean "cannot fork". The source's backend session id must already be pinned (let it run a turn first). The fork inherits the source's repo + backend; --type

defaults to development

. Unlike warden review

/ warden models

, fork is a managed spawn that crosses the daemon, so it has MCP + CLI parity — the fork_agent

MCP tool is the orchestrator twin (a thin wrapper over spawn_agent

with fork_from

set, no new endpoint).

Manage the bearer token that gates non-loopback (remote) access — see Remote access.

warden token generate                 # mint a 256-bit token (persisted to ~/.warden/token.env, 0600)
warden token show                     # print the current token (paste into a remote client)
warden token rotate                   # regenerate in place + restart the daemon
warden token show --readonly          # print the read-only token (WARDEN_READONLY_TOKEN)

The WARDEN_TOKEN

env var overrides the file so the secret can stay off disk.

Read-only token. Set an optional second token, WARDEN_READONLY_TOKEN

, to hand out view-only access (e.g. a shared dashboard or a teammate). A read-only token may read everything — all GETs plus the live event stream — but is denied every state-changing action and the interactive attach. Mint one with warden token generate

and export it as WARDEN_READONLY_TOKEN

(it only works alongside a primary WARDEN_TOKEN

; the daemon refuses to start with a read-only token but no primary one).

warden's interactive mode: a proper terminal REPL to drive the fleet, with a real line editor (arrow keys, persisted history, reverse-search, a live /-command menu that filters as you type,

Tab completion, colourised prompt) that closes cleanly with Ctrl-D. It drives the fleet two ways:

Deterministic/

commands (no model)/agents

,/spawn <prompt>

,/tell <id> <text>

,/memory <question>

,/pipelines

, … Typing/

pops a live, filtering menu of matching verbs (each with its summary);/help

lists them all. These keep working even when the local model is slow or wrong. When a command needs more input, aguided argument form collects it — numbered pick-lists for known fields (model, permission, type), free text for the rest — opening automatically for a missing required arg or on a+

-suffixed verb (/spawn+

); a local model, if present, pre-fills each field with a suggestion you can accept, override, or clear.Natural language (local LLM)— any other line is planned into** confirmed**warden tool calls without spending cloud-model tokens. It conducts; it never implements — all code work is delegated by spawning an agent.Local project grounding (— ask a project question ("where does the spawn gate live?", "how do I run the tests?") via$0

)/memory <q>

(/mem

//ask

) or theproject_memory

tool and warden answers itlocally from.warden/memory.md

, citing each entry's trust + provenance. Itremovesa cloud round-trip (rather than adding tokens like injection); read-only, degrades to the matching entries verbatim with no local model, default on viamemory.ground

.

It starts without a local model (the /

commands and !

-shell always work); only the natural-language half needs local_llm: true

. Every mutating action passes a mandatory confirm gate. Run standalone, or as the cockpit master pane via the repl

config / --repl

flag (Alt+t toggles it with a raw shell). See docs/FEATURES.md §17.

warden repl                           # aliases: warden interactive, warden i

Inspect the plugin registry — external executables that extend warden with custom agent task types and lifecycle hooks (over a versioned JSON-over-stdio protocol). Default off (plugins: true

to enable, since plugins run external code).

warden plugin list                    # registered plugins: paths, custom task types, subscribed hook events, config errors

Hooks (pre/post-spawn

, pre/post-commit

, pre/post-check

) are advisory and fail-open — a missing, slow, or crashing plugin is logged and skipped, never blocking an agent. Configure via plugins.enabled

  • a plugins.registry

list in ~/.warden/config.yaml

; a worked example lives under examples/plugins/. See

docs/FEATURES.md §26.

Run the daemon (HTTP API + background poller). Normally managed by launchd (macOS) or systemd (Linux); run manually for debugging.

warden daemon
warden daemon --addr 127.0.0.1:9000

Run the MCP stdio server so an orchestrator agent session (e.g. Claude) can manage agents via tool calls.

warden mcp
warden mcp --addr 127.0.0.1:8765

Tools exposed: list_agents

, get_agent

, spawn_agent

, adopt_agent

, send_to_agent

, get_agent_output

, stop_agent

, terminate_agent

, restore_agent

, delete_agent

, remove_worktree

, ctx_set

, ctx_get

, ctx_list

, ctx_cas

, ctx_append

, send_message

, read_inbox

, wait_for_message

, list_approvals

, approve

, commit

, push

, sync

, check

, get_collaboration_status

, who_is_editing_file

, get_branch_status

, create_pipeline

, start_pipeline

, show_pipeline

, list_pipelines

, cancel_pipeline

, list_schedules

, snapshot_create

, snapshot_list

, snapshot_restore

, insights

, savings

, spend

.

Generate shell completion scripts for Bash, Zsh, Fish, or PowerShell.

warden completion bash   # generate bash completion script
warden completion zsh    # generate zsh completion script
warden completion fish   # generate fish completion script
warden completion powershell   # generate PowerShell completion script

The completion script should be redirected to the appropriate location for your shell. See the Shell Completion section for installation instructions.

Register warden mcp

as an MCP server in your orchestrator agent's MCP config. For a Claude Code orchestrator that's ~/.claude/claude_desktop_config.json

or the project-level .claude/mcp.json

; other MCP-capable agents use their own config path:

{
  "mcpServers": {
    "warden": {
      "command": "warden",
      "args": ["mcp"]
    }
  }
}

warden mcp

connects to the daemon at the addr

config setting (default 127.0.0.1:8765

); to point it elsewhere use "args": ["mcp", "--addr", "host:port"]

.

Once registered, the orchestrator session can call these tools directly:

Tool Description
list_agents
List all active agents with their status, working directory, and subject
get_agent
Get full detail (status, workdir, subject, events, worktree) for one agent
spawn_agent
Spawn a new agent — pass a prompt for a quick auto-typed agent, or type +repo for a managed worktree; set supervised: true for --permission-mode acceptEdits instead of full bypass; pass role to attach a built-in role persona + its default flags
set_role / list_roles
Switch a running agent's built-in role (relaunches to re-inject the persona; general /empty clears it) / list the fixed built-in role catalog (name + description) for a picker
adopt_agent
Register an existing Claude session: resume newest-for-dir under tmux, or live-register a running tmux session
send_to_agent
Type a message into a specific agent's claude session
get_agent_output
Return the recent terminal output of a specific agent
stop_agent
Umbrella teardown. Default = full teardown (terminate + clear record + remove worktree). keep_record / keep_worktree subtract steps (keep_worktree alone == the old done ); hard purges the record; pr /base open a GitHub PR first while the agent is intact; force /delete_adopted_branch for the worktree guards. Destructive when it removes the worktree — only after explicit user confirmation
terminate_agent
Stop an agent (kill tmux + the agent process); keeps the record and worktree. Reversible via restore_agent — the default "stop this agent" action
restore_agent
Recreate and resume a lost/orphaned agent's session (claude --resume )
recover_agents
Safety net for the tombstone reaper: revive archived records whose tmux session is confirmed still alive. apply:false (default) only reports candidates; apply:true re-inserts each one under its original id, reconnecting any children automatically
delete_agent
Clear an agent's stored record (archives by default; hard purges). Does not touch tmux or the worktree
remove_worktree
Remove an agent's git worktree + branch — destructive; refuses while the agent runs or has uncommitted/unpushed work unless force
ctx_set / ctx_get / ctx_list / ctx_cas / ctx_append
Read/write the shared-context key/value blackboard agents collaborate through (with compare-and-swap and append)
send_message / read_inbox / wait_for_message
Send a directed message to an agent (wakes it if parked) / read this agent's inbox / block until a message arrives
list_approvals / approve
List recognized pending tool-permission prompts / answer one by option number
commit / push / sync
Git lifecycle on the agent's pinned worktree (staged commit with auto-message, push — force uses --force-with-lease , rebase-sync) returning compact structs instead of raw git output
check
Run the project's .warden/check.yml checks, returning pass/fail with output for only the failing ones
get_collaboration_status / who_is_editing_file
File-conflict view across the fleet / who (if anyone) is editing a given file
create_pipeline / start_pipeline / show_pipeline / list_pipelines / cancel_pipeline
Drive a DAG pipeline — create from a YAML spec, start entry jobs, inspect, list, cancel

Pipeline MCP tools are thin wrappers over the same daemon routes the CLI uses, so an orchestrator session can run a multi-stage workflow (analyze→implement→review) without shelling out. Lifecycle controls (

_pipeline

,resume_pipeline

,delete_pipeline

,edit_pipeline_job

,retry_pipeline_job

) are available over MCP too, mirroring thewarden pipeline …

CLI.

Example orchestrator prompts:

  • "What is PROJ-350 doing?" — calls get_agent

to fetch current status and events - "Tell PROJ-343 to run the tests" — calls send_to_agent

with"run the tests"

  • "List all my agents" — calls list_agents

  • "Spin up an agent to research SSE reconnection" — calls spawn_agent

with aprompt

(auto-typed) - "Spawn a debug-ci agent in /path/to/repo" — calls spawn_agent

withtype

+repo

  • "Stop PROJ-350" — calls terminate_agent

(reversible); "clear its record too" — thendelete_agent

Beyond raw tool access, install the packaged Claude Code skill so a Claude orchestrator session knows how and when to manage your fleet (triage, create-from-prompt, relay "tell X to do Y", terminate-with-confirmation, daemon-down handling). Other MCP-capable orchestrators drive the same tools directly without the skill:

make install-skill   # symlinks skills/warden into ~/.claude/skills/warden

With the MCP server registered (above) and the skill installed, just talk to a Claude session: "list my agents", "spin up an agent to research X", "what is agent-4f2a doing?", "tell agent-4f2a to run the tests", "kill the idle ones" — it drives the MCP tools (falling back to the warden

CLI if the MCP server isn't registered). The daemon must be running.

warden moves deterministic responsibilities off agents — git and checks — onto the first-class warden commit

/push

/sync

/check

commands (CLI + MCP), and enforces the worktree boundary with PreToolUse hooks delivered through a per-agent claude --settings

file (Claude Code backend). Each hook fails open (a hook error never blocks the agent) and is individually config-gated (default on):

Layer Setting What it does
Prompt steer
git_conventions
A system-prompt hint steering agents toward wd commit /push /sync /check over raw git/test Bash — the gentle first layer
Isolation guard
isolation_guard
Denies an isolated agent's Edit/Write that escapes its worktree into the shared repo
Git-guard
git_redirect
Deny-redirects raw git commit /push /pull /rebase to the warden tools (reads stay allowed)
Check-guard
check_redirect
Deny-redirects a raw test/lint/build command registered in .warden/check.yml to wd check (focused -run invocations pass through)

The default write-type isolation (each code

/docs

/website

/debug-ci

/tests

agent gets its own worktree unless --in-repo

) is what makes the isolation guard meaningful and fixes parallel-agent collisions. See docs/FEATURES.md §22 for the full model.

./scripts/install.sh   # install + start as a background launchd service (recommended)
make run-daemon        # foreground, for debugging only (blocks the terminal; ctrl-C to stop)

warden start PROJ-350 --type development

warden ls
warden status PROJ-350

warden attach PROJ-350

warden done PROJ-350
make build            # go build -o bin/warden ./cmd/warden
make test             # go test ./...  (includes fuzz seed corpora)
make lint             # go vet ./...
make run-daemon       # build + start daemon in the foreground (debugging only)
make test-integration # build-tagged end-to-end suite (real daemon subprocess + CLI)
make bench            # run the Benchmark* suite (store I/O, pipeline, approvals)
make fuzz             # deeper fuzz sweep of the YAML/pane/session parsers
make cover            # whole-repo statement coverage (prints the total%)

The README coverage badge is a static snapshot of make cover

's total — refresh it when coverage moves materially.

All unit tests run without Docker or any external services:

go test ./...

The integration suite (make test-integration

, build tag integration

) boots a real warden daemon

against an isolated HOME

; its spawn-lifecycle test self-skips unless tmux

and claude

are installed, so it stays CI-safe.

The repo ships version-controlled hooks in .githooks/

. Wire them once with:

make install-hooks          # git config core.hooksPath .githooks

scripts/install.sh

(the service installer) also does this automatically. Once wired:

pre-commit runsmake fmt-check lint

(gofmt +go vet

) — fast, so a formatting/vet slip is caught before the commit lands.pre-push runsmake verify-fast

(gofmt/vet/web/release build) — the CI subset that doesn't need an isolated machine.

Bypass either in a pinch with git commit --no-verify

/ git push --no-verify

. Because core.hooksPath

is a relative path, the hooks work across git worktrees too.

The daemon embeds a React dashboard (Astro + React) and serves it at http://localhost:8765

alongside the REST API — no separate server required.

make release     # 1. builds the Astro UI (web/), 2. embeds it via go:embed, 3. builds bin/warden
warden daemon  # start the daemon as usual

Then open http://localhost:8765

in a browser.

Note:the UI is baked into the binary at build time. After changing anything underweb/

, re-runmake release

(ormake ui

to rebuild only the frontend) and restart the daemon for the changes to take effect.

The dashboard is a routed mission-control shell. Tabs are real URLs (History-API routing — back/forward, refresh, and shareable deep links all work), so each surface has its own address:

Route Tab What's there
/cockpit
⊞ Cockpit
The home view (/ redirects here). A slim Fleet header (totals · busy · waiting · errored, pressure, dirs) above the full agent grid.
/others
▦ Others
The former Overview, now a catch-all: Needs you (attention queue), File conflicts, and Recent activity.
/pipelines
⛓ Pipelines
Pipeline list + live DAG / job drawer.
/metrics
📊 Metrics
Per-agent and fleet-wide charts — see below.
/archive
🗄 Archive
Ended sessions with since/type filters.
/agent/<id>
<id>
A pinned agent's live terminal (one closeable tab per pinned agent).

Cockpit is the home/

redirects to/cockpit

. It carries theFleet summary header (moved out of the old Overview) above the canonical agent grid; the redundantQuick spawnwidget and the duplicateAll agentsmini-grid were removed.Metrics tab (— a scrollable column of uPlot chart cards:/metrics

)CPU per agent,** Memory per agent**(GiB),** Cost per agent**(live measured model spend in $, with a total/today/this-week headline and a sorted per-agent cost bar chart — top-N costliest, rest folded into anothers

row),Context per agent(a client-accumulated time series of each agent's live context fill, legend dot colored byok

/warning

/critical

; in-session only — resets on full reload),Number of agents(fleet size over time), and** Tokens saved**(daily bars from the savings ledger + a headline saved-tokens/$ figure). When the savings ledger is disabled the savings/cost cards show a "setsavings: true

" hint instead of an empty chart. ALive footprint card carries the former Resources panel.Context & Messages— no longer a tab; opened from a small🗒 button in the header as a dismissible overlay (Esc to close).Agent tabs— pin any agent to its own tab to get a** live, interactive terminal**(AttachTerminal

) — a realtmux attach

bridged to the browser over a WebSocket, so you can type into the agent and watch it respond in real time.Create agent+ New agent opens a prompt box (with a directory picker and aSupervised checkbox). Type the task and pressCreate(or Cmd/Ctrl+Enter); the type label is assigned automatically. Tick** Supervised**to launch with--permission-mode acceptEdits

instead of full bypass. For a managed worktree, use the CLI:warden start TICKET --type development --repo …

.Terminate— surfaces the git guard (409 →** Force**+ optional** hard-delete**) when there's uncommitted/unpushed work.** Agent grouping**— the Cockpit grid buckets agents into collapsible panes by** Directory / Type / Status / Tag**(choice saved to LocalStorage).** Batch operations**— per-tile checkboxes (Shift-click range select) raise a bulk action bar offering** Message…, Terminate**, and** Deleteacross the selection. Search & Archive**— a live search box filters the fleet client-side; the 🗄** Archivetab browses ended sessions with since/type filters. Theme toggle**— header control cycles** System → Light → Dark**(defaults to System; persisted, applied before first paint).** Keyboard shortcuts**— a global layer:?

help overlay,n

new agent,/

focus filter,r

refresh,1

9

jump to a tab,j

/k

next/previous tab,Esc

close/blur (also closes the Context overlay).Browser notifications— opt in to get a desktop notification when an agent enterswaiting_for_input

(gated so they only fire while the tab is hidden).

The dashboard is mobile-responsive and can be reached from your phone or another machine. Generate a token (export WARDEN_TOKEN=$(warden token generate)

), bind a non-loopback address (warden daemon --addr 0.0.0.0:8765

), and expose it over Tailscale (recommended) or a Cloudflare Tunnel. The daemon refuses to bind non-loopback without a token; the browser prompts for it on first load. See docs/USAGE.md → Remote access for the full walkthrough.

Run two terminals in parallel — no rebuild loop needed while iterating on the UI:

warden daemon

make ui-dev

Open http://localhost:4321

. Edits under web/src/

trigger HMR instantly; the browser stays on the same origin as the real daemon API so SSE and all REST calls work without CORS configuration.

make web-test    # Vitest — frontend unit tests (status mapping, API client)
go test ./...    # Go suite — covers daemon hub, SSE endpoint, static embed, and all existing routes

The frontend Vitest suite lives in web/src/lib/

alongside the source files (status.test.ts

, api.test.ts

). The Go daemon tests cover the broadcaster (hub_test.go

), the SSE handler (sse_test.go

), and the static file serving with SPA fallback (static_test.go

).

Issues and pull requests are welcome. See CONTRIBUTING.md for licensing, the DCO sign-off (git commit -s

), and dependency/notices rules. Before opening a PR:

gofmt -l $(git ls-files '*.go')   # must be empty (CI enforces gofmt)
make lint                          # go vet ./...
make test                          # go test ./...
make web-test                      # frontend unit tests

CI (build, test, lint) runs on every push and PR to main

— see .github/workflows/ci.yml.

Licensed under the Apache License, Version 2.0. See LICENSE and NOTICE. Third-party dependency licenses are listed in THIRD-PARTY-NOTICES.md.

warden is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, Google, Cursor, Block, or any other agent-backend vendor. "Claude" and "Claude Code" are trademarks of Anthropic, PBC; "Codex" of OpenAI; and all other product and company names are trademarks of their respective owners, used here for identification only.

warden is a local orchestrator: it drives agent-backend CLIs and services using credentials you supply. You are responsible for complying with the terms, usage policies, and rate limits of every backend and account you connect. warden grants no rights to any third-party service and makes no representation that a given automation or concurrency pattern is permitted under your plan.

Copyright 2026 Srajan Pathak

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0
── more in #developer-tools 4 stories · sorted by recency
── more on @warden 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/warden-run-a-fleet-o…] indexed:0 read:74min 2026-07-27 ·