What if heterogenous agents could talk to each other?

Repowire, a new open-source tool, connects multiple coding agents like Claude Code and Codex across different repositories, allowing them to pass messages and coordinate tasks without manual copy-paste. The system runs locally as a daemon on macOS or Linux and provides a control layer for multi-agent work, enabling agent-to-agent questions, scheduled reminders, and remote monitoring through a browser dashboard, Telegram, or Slack. Repowire aims to solve the problem of isolated AI coding assistants by creating a mesh where agents can communicate, dispatch work, and track durable jobs across sessions.

Repowire connects the coding agents you already have open. Claude Code in one repo, Codex in another, a dashboard in your browser, Telegram on your phone: Repowire gives them names and lets them pass messages without copy-paste. It is a local control layer for multi-agent work: ask another agent a question, send a quick update, schedule a reminder, or run one session as the coordinator. Use it when: - One repo needs a concrete answer from an agent already working in another repo. - You want a personal orchestrator session to dispatch tasks, collect status, or keep reviews moving. - You want to monitor or nudge agent work from your phone or browser. - A session should wake itself or another peer later with a scheduled check-in. Repowire runs locally by default through a daemon on your machine. The hosted relay is optional and uses outbound connections for remote dashboard access and cross-machine mesh traffic. Requirements: macOS or Linux, Python 3.10+, tmux. 1. Install Repowire and wire your agents. curl -sSf https://raw.githubusercontent.com/prassanna-ravishankar/repowire/main/install.sh | sh repowire setup The installer detects uv , pipx , and pip in that order. 2. Open your normal agent CLIs. Use the tools directly; Repowire hooks into them after setup. tmux window 1 cd ~/projects/project-a && claude tmux window 2 cd ~/projects/project-b && codex 3. Check that both peers appeared. Claude Code registers on session start. Codex registers after its first interaction, so send a short warmup prompt in project-b , then run: repowire peer list 4. Ask from one agent to the other. In project-a , tell your local agent: Ask project-b what API endpoints they expose. Your local agent invokes Repowire's ask MCP tool, the second agent receives the question, and the reply comes back as an ack notification. Repowire is the mesh and tool surface around the agents, not a standalone chat UI. The same pattern works across Claude Code, Codex, Gemini CLI, OpenCode, and Pi when those runtimes are installed. multi-repo-fast.mp4 You can also spawn peers through Repowire: repowire peer new ~/projects/project-a repowire peer new ~/projects/project-b --backend codex --profile fast For durable recurring workers, scaffold a repo-local agent folder and target it from jobs: repowire agents create daily-brief --backend codex repowire jobs create "Daily brief" --path .repowire/agents/daily-brief --backend codex --cron "@daily" --prompt "Prepare the brief." Full docs: docs.repowire.io https://docs.repowire.io . Agent-to-agent asks : Non-blocking questions with explicit ack replies and reminder injection until a thread is closed. Human control surfaces : Browser dashboard, Telegram, and Slack can route messages as service peers. Durable jobs : Track one-off and recurring work through CLI/MCP, with dashboard visibility and controls for run, retry, and cancel. Orchestrator pattern : A dedicated peer can dispatch work, check status, coordinate reviews, and keep a queue moving. Scheduled wake-ups : Send a future notification or ask to yourself, another peer, or an orchestrator. Optional relay : Reach the dashboard remotely and bridge machines without opening inbound ports. All peers connect to a local daemon. The daemon keeps the registry, routes asks/notifies, tracks open asks, stores durable jobs, runs schedules, and feeds the dashboard timeline. The stable public surface is peers, circles, asks, notifications, broadcasts, schedules, and the /jobs tracked-work API exposed through CLI and MCP JSON tools. The v0.14 direction is session-native: sessions become the durable unit of work, while peers remain the live runtime executors. The current dashboard shows the selected peer/session view, merges Claude transcript history where available with realtime events, and is moving toward broader session commands for controls like resume, scheduling, approvals, and future backend/model changes. Transport notes: - Claude Code, Codex, and Gemini CLI use hooks plus MCP tools. - OpenCode uses a TypeScript plugin plus WebSocket. - Pi uses the Repowire extension path when detected by setup. - Claude Code channel/ACP delivery is experimental and opt-in. - Relay is optional remote access, not a requirement for local routing. | Agent runtime | Connection path | |---|---| | Claude Code | Hooks + MCP; optional experimental channel/ACP transport | | Codex | Hooks + MCP | | Gemini CLI | Hooks + MCP through normalized BeforeAgent / AfterAgent events | Antigravity CLI agy | Plugin install verified; hook firing and MCP pending upstream verification | | OpenCode | Plugin + WebSocket | | Pi | Repowire extension | | Human or service surface | Role in the mesh | |---|---| | Dashboard | Browser control surface at localhost:8377/dashboard or through relay | | Telegram | Phone control surface and notification target | | Slack | Team chat control surface | | Orchestrator peer | Long-running coordinator that dispatches and reviews work | | Relay dashboard | Optional remote dashboard and cross-machine bridge | repowire setup auto-detects installed runtimes and wires the supported transports it finds. The agent-orchestration space is moving fast. Most projects cluster around a few shapes: Worktree/task runners : Claude Squad https://github.com/smtg-ai/claude-squad , Vibe Kanban https://github.com/BloopAI/vibe-kanban , and dmux https://dmux.ai/ help launch and review many isolated agent workspaces. Deterministic schedulers : Bernstein https://github.com/sipyourdrink-ltd/bernstein decomposes goals, runs agents in parallel worktrees, verifies, and merges passing work. Hierarchical swarms : multi-agent-shogun https://github.com/yohey-w/multi-agent-shogun defines manager/worker roles and routes tasks through tmux, files, or role-specific protocols. Agent IDEs and workflow systems : HumanLayer/CodeLayer https://github.com/humanlayer/humanlayer focuses on planning, review, team workflows, and richer agent workspaces. Repowire sits in a different slot: it is a live mesh and control plane for agent sessions you already have running. It does not try to be the scheduler that decomposes every goal, the kanban board that owns every branch, or the merge gate that lands code. It gives your existing terminals, dashboard, Telegram, Slack, and orchestrator session a shared address book, message lifecycle, schedule queue, and local session timeline. Project A needs the real API shape from Project B. Ask project-b ; the peer answers from its live checkout, not stale docs. See multi-repo coordination https://docs.repowire.io/use/workflows/multi-repo/ . Send work to a peer from Telegram, Slack, or the dashboard, then receive progress updates from agents as notifications. Telegram and Slack human messages open tracked asks by default; use their notify/FYI commands for fire-and-forget nudges. See mobile mesh management https://docs.repowire.io/use/workflows/mobile-mesh/ . Run one session as the orchestrator. It can dispatch to project peers, ask for status, review PRs, and wake itself later. See orchestrator coordination https://docs.repowire.io/use/workflows/orchestrator-coordination/ . Schedule a reminder, check-in, or future ask: repowire schedule self 10m "check CI" repowire schedule create orchestrator 1h "handoff" --from-peer project-a --kind ask Enable the hosted relay when you want remote dashboard access or cross-machine mesh traffic: repowire setup --relay The dashboard shows peers, status, descriptions, chat turns, tool calls, attachments, durable jobs, and the selected peer/session timeline. For Claude Code peers, it can merge transcript history with realtime events; other backends contribute realtime events as their transports report them. Run it locally at: http://localhost:8377/dashboard With relay enabled, use: https://repowire.io/dashboard repowire setup install hooks/MCP/plugin/service for detected agents repowire setup --http-mcp opt in to localhost Streamable HTTP MCP at /mcp repowire setup --update-checks let status/doctor report available updates repowire update explicit package upgrade + hook reinstall + daemon restart repowire status show installed components and daemon status repowire doctor run diagnostics repowire service restart restart the installed daemon service repowire peer list list mesh peers repowire peer new PATH --profile P spawn a peer in tmux repowire schedule self 10m "check CI" wake this peer later repowire telegram start run Telegram service peer repowire slack start run Slack service peer The daemon uses ~/.repowire/state.db for durable local state. On first startup after install or update, it applies SQLite migrations and imports legacy schedules.json , events.json , and sessions.json once while leaving those files in place for downgrade/export compatibility. Migrated state is written to SQLite, and repowire doctor reports the SQLite schema, integrity, and import status. See the full CLI reference https://docs.repowire.io/reference/cli/ and MCP tools reference https://docs.repowire.io/reference/mcp-tools/ . Config lives at ~/.repowire/config.yaml . daemon: host: "127.0.0.1" port: 8377 auth token: "rw local ..." mcp http: enabled: false bind: "localhost-only" require auth: true allow dangerous tools: false spawn: commands: claude-code: "claude --dangerously-skip-permissions" codex: "codex --dangerously-bypass-approvals-and-sandbox" gemini: "gemini --yolo" profiles: codex: fast: args: "--model", "gpt-5-mini" allowed paths: ~/git, ~/projects updates: check enabled: false relay: enabled: true url: "wss://repowire.io" api key: "rw ..." Update checks are off by default. If enabled with repowire setup --update-checks , repowire status and repowire doctor may report that a newer release is available, but they do not upgrade packages, rewrite hooks, or restart services. Use repowire update when you want to upgrade explicitly; it preserves enabled package extras such as repowire acp where practical. Security defaults: - Local daemon binds to 127.0.0.1 . - Relay is opt-in and uses outbound WebSocket. - WebSocket and local HTTP auth are available through daemon.auth token . - Experimental HTTP MCP is opt-in, localhost-only, bearer-authenticated by default, and not exposed through the hosted relay. - Spawn requires explicit command and path allowlists. - Experimental channel/ACP transport is opt-in. git clone https://github.com/prassanna-ravishankar/repowire cd repowire uv sync --extra dev uv tool install . --force-reinstall Hooks and MCP servers run the installed repowire executable, not your checkout. After changing daemon, hook, or MCP code locally, reinstall the tool and restart the daemon service so the live mesh uses the new code: uv tool install . --force-reinstall repowire setup --non-interactive rewrites hooks/MCP/service to the installed local build repowire service restart enough when only daemon code changed If service management fails, use repowire service status first. Raw launchctl on macOS or systemctl --user on Linux are fallback troubleshooting tools. repowire uninstall uv tool uninstall repowire repowire uninstall removes hooks, MCP entries, channel transport config, OpenCode plugin files, and the daemon service. It does not automatically remove ~/.repowire/ , which contains local config, events, attachments, and relay keys. See CONTRIBUTING.md /prassanna-ravishankar/repowire/blob/main/CONTRIBUTING.md . Before opening a PR, run the advisory repo-hygiene checklist: python3 scripts/pre pr hygiene.py It is an opt-in prompt for docs, README, agent-instruction, and graphify follow-ups, not a mandatory hook. It also flags Beads JSONL ledger churn before it can leak into PR diffs. MIT