cd /news/ai-agents/i-built-a-harness-and-now-it-is-my-d… · home topics ai-agents article
[ARTICLE · art-127910] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

I built a harness and now it is my daily driver for local agents

Developer mrsirg97-rgb released rig, a single-binary local agent runtime that assembles context, streams an OpenAI-compatible model endpoint, executes tool calls, and repeats, shipping 18 built-in tools and a Python plugin system. rig requires Go 1.26.6 or later for the go install path, defaults its endpoint to http://127.0.0.1:8090/v1, and has no model default — a run without one refuses at start. Configuration lives in ~/.rig/ across settings.json, models.json, AGENTS.md, theme.json, and plugins/, with each key resolving in the order flag, environment, file, built-in default.

read5 min views1 publishedSep 12, 2026
I built a harness and now it is my daily driver for local agents
Image: Michielbdejong (auto-discovered)

A minimum runtime for your agents.

One binary. One model endpoint. One terminal.

rig assembles context, streams the model, executes tool calls, returns results, and repeats. The TUI, piped CLI, headless worker, and dashboard share the same session, task, memory, and scheduler stores.

Choose one:

Installer (POSIX sh, no Go, no sudo; installs to ~/.local/bin):

curl -fsSL https://mrsirg97-rgb.github.io/rig/install.sh | sh

Release binary from releases/latest. Choose your <os>_<arch>:

curl -fsSL https://github.com/mrsirg97-rgb/rig/releases/latest/download/rig_linux_amd64 -o rig
chmod +x rig

go install (needs Go ≥ 1.26.6; the core is stdlib-only):

go install github.com/mrsirg97-rgb/rig/cmd/rig@latest

rig -update fetches, verifies, and atomically installs the latest release. The running process keeps the old binary until restart.

./rig --base-url $ENDPOINT --model $NAME

rig needs an OpenAI-compatible SSE endpoint and a model ID. The endpoint defaults to http://127.0.0.1:8090/v1; there is no model default — a run without one refuses at start, naming the three ways to set it (--model, RIG_MODEL, the model key in settings.json). The TUI is the frontend when stdout is a terminal, the piped CLI otherwise. For scripts, run ./rig -p "the task". See docs/SETUP.md for configuration.

rig ships 18 built-in tools. Restrict them with --allow:

tool what it does
bash run shell commands; output bounded
read /write /edit files; edits are exact-match, provenance-checked
ls /find /grep the filesystem, by name and by content
diff the working tree against HEAD, or a tool's two latest observations
python a persistent IPython kernel; variables and imports survive
web_search a local SearXNG instance
web_fetch a URL as readable text; private addresses refused
todo the task queue, scoped to the project (a repo's worktrees share one)
rem memory across sessions: learn, recall, reflect, prune; scoped to the project
scheduler background jobs on your crontab, run in a bubblewrap jail
delegate a one-shot headless worker for a bounded subtask
sessions read-only vitals of the session store
plugin /plugins the door into your python plugins, and their ecosystem

Every tool result is capped. Repeated identical failures are bounded. An optional round cap limits calls per turn. A failed call executes once.

A Python plugin is one file and one tool. It provides run and schema. There is no build step. Model-authored plugins land in ~/.rig/plugins/pending/. Approve, disable, and reload them with /plugins or the dashboard. See docs/PLUGINS.md.

Configuration lives in ~/.rig/. Set $RIG_HOME to move it. Every file is optional. Invalid files fail startup and name the file and field.

file what it holds
settings.json the knobs: endpoint, model, the allow-list, the retry bound, the approval dial, the worker sandbox
models.json the per-model table: context window, max tokens, the compaction reserve, the role ( worker /interactive ), the effort levels
AGENTS.md global instructions, read before the project's <cwd>/AGENTS.md
theme.json the terminal theme: base, slot colors, glyph set
plugins/ your python plugins (top-level files are live)

Each key resolves in this order: flag, environment, file, built-in default. /models lists and switches models. /effort changes reasoning effort. See docs/SETUP.md for configuration and sandbox settings.

rig serve

The dashboard serves the rig stores on loopback only. On first run it prints an access token, stores it with mode 0600, and includes it in the URL. The page exchanges the token for a cookie. Mobile friendly.

  • sessions : list them per workspace, and resume one mid-work
  • todo : the queue, with create, start, complete, and retry
  • scheduler : the jobs, with create, , resume, remove, an in-place update form that opens with the job's current fields, and each job's run audit trail
  • models : the table, with the effort dial
  • plugins : approved, pending, disabled; the forge reads and saves a plugin's source into the pending zone
doc what it is
docs/DESIGN.md architecture, the seams, turn semantics, extension guide
docs/SETUP.md build, configuration, verification
docs/USAGE.md running a session; session and failure semantics
docs/PLUGINS.md the python plugins: the contract, the zones, creating and consuming
SECURITY.md the trust model and how to report a vulnerability
CONTRIBUTING.md the process: spec first, tests before code, the freeze
cmd/rig      composition root; wires every seam once; flags and env only
core            the seams, wire types, and the streaming-event vocabulary
loop            the concrete turn runtime (fault/cancel-aware)
evt             the event loop (SPEC_EVT): one consumer, many producers; the
                turn runtime's engine
kernel.go       the composition kernel
command/        the user commands (/compact, /models, /sessions, /effort, ...)
config/         the four-layer config resolution (flag > env > file > embedded)
models/         the per-model table (window, compaction numbers, role, effort)
policy/         ContextPolicy implementations: compact (per-model trigger),
                effort (the reasoning dial's provider decorator)
middleware/     ToolMiddleware: toolset (the live table), approve (the gate),
                paths (the ~ boundary), perm (deny by default + plugin
                provenance), guard (the bound, the round cap, the result cap)
provider/       Provider implementations (the openai-compatible SSE adapter)
plugins/        python plugin discovery (one file, one tool) and the plugin
                door (run/schema) and the ecosystem (list/create/delete/reload)
store/          the SQLite stores (state, todo, rem, scheduler), the sqlx
                transaction seam, the project scope identity (store/scope);
                -resume projects a session back from the state rows
tool/           Tool implementations: bash(1); file read/write/edit; fs
                ls/find/grep; todo the job queue; rem memory; scheduler
                background jobs; delegate the one-shot worker; python the
                persistent IPython kernel; web search and fetch; diff the
                observation diff; sessions the soak's vitals
frontend/       Frontend implementations: cli (the piped reference), tui (the
                terminal default), oneshot (-p worker), web (the serve
                dashboard)
specs/          the specs, written and agreed before the code (SPEC_CORE first)
docs/           DESIGN (architecture), SETUP (build/config), USAGE (running),
                PLUGINS (the python plugins)

The structural test is simple: add one file and one registration line. The loop never names a concrete tool, provider, policy, frontend, or middleware. A Python plugin needs no Go. See docs/DESIGN.md, docs/PLUGINS.md, and CONTRIBUTING.md for the process.

core/ and loop/ use only the standard library. Stores use the pure-Go modernc.org/sqlite driver.

── more in #ai-agents 4 stories · sorted by recency
── more on @rig 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/i-built-a-harness-an…] indexed:0 read:5min 2026-09-12 ·