Lanchu

What problem does it solve?

More and more people use several AI agents (like Claude or Cursor) to build apps or automate their company. But when you put several agents to work at once, two pains show up:

They don't coordinate — and they get in each other's way:

• They step on each other's work — two agents do the same thing.

• They work blind — one doesn't know what the other did.

You can't control them — or trust them:

• They stray out of their lane — an agent touches something it shouldn't.

• The documentation goes stale — nobody keeps the knowledge up to date.

• You can't see anything — you don't know who did what, or what they spent.

Most tools only tackle coordination, and they're built for programmers. Lanchu adds what's missing: control and trust for whoever supervises.

Lanchu does not orchestrate your agents (it doesn't decide their plan). It gives them a shared workspace so they coordinate without colliding, sets scope limits on them, keeps the documentation up to date, and gives you a panel + history to see and trust what they do — even if you're not technical.

Related MCP server: MCP Orchestration Server

How it works (the idea)

1. You launch your agents as always (with the tool you already use).

2. Each agent registers in your organization and takes on a role.

3. From there, it only works on what it's responsible for: it claims tasks, reads the shared documentation, and Lanchu rejects and records any action outside its lane. The agents coordinate through Lanchu, not by talking to each other — that's why you can see and bound everything.

4. You watch the real-time panel: who's active, what they're working on, what documentation they create, and a history of everything they did.

Lanchu sets cooperative, auditable limits: it blocks what passes through it and leaves everything in plain sight. It's not a system cage — the trust comes from seeing it all.

How it fits

Agents coordinate through Lanchu (a shared blackboard), never directly with each other — so every action is visible and can be bounded.

flowchart TB
    A1["Agent A"] -->|MCP| L
    A2["Agent B"] -->|MCP| L
    subgraph L["Lanchu — local server"]
        S["Shared state · roles · audit log"]
    end
    L --> P["Supervisor panel<br/>(real-time + history)"]
    L -. resource updates .-> A1
    L -. resource updates .-> A2

Quick start

npx lanchu "fix the login"

Note: Requires Node >= 22.5. See CHANGELOG.md for what's new, DEFINITION.md for the full picture and CLI.md for the command surface.

What's included today

Coordination

• Organizations, projects, roles and tasks — a shared blackboard with scope control; actions outside the role are rejected and recorded.

• Agent isolation — lanchu spawn gives each agent its own terminal, git worktree and branch, so parallel work can't collide. lanchu tile arranges the terminals into a mosaic.

• Agent-to-agent messaging with conflict warnings — audited, never private; idle agents with queued notices are woken automatically.

• SDLC pipeline (assist mode) — the server owns the lanes (definition → build → review → qa → done): attach a PR and the task routes to review; finish it and it routes to QA verification. Agents can bounce underspecified tasks back to definition (task_reject), with the reason audited.

Governance

• Roles you can edit — tag scopes, wildcard, a per-role token quota (claims warn at 80%, block at 100%) and a preferred model tier; lanchu spawn --model overrides per spawn.

• Coordinator lease — one coordinating agent per org, enforced; the supervisor grants or revokes it with lanchu coordinator.

• Greenzone restarts — lanchu restart --greenzone coordinates a maintenance window: every agent confirms a safe point before the server goes down.

• Session security — per-session Bearer tokens (never in window titles or ps args); lanchu rotate-tokens invalidates every open session after an exposure.

Knowledge & memory

• Shared, traceable documentation — categorized docs with section and delta reads, read tracking and usage analytics.

• Persistent memory — three scopes (agent / project / org) stored in the org DB and auto-distilled into each agent's context across sessions.

Observability

• Real-time panel — Overview home (who's working right now), Projects, Team, Work board, Org life (a force-directed graph of the org built from the audit log), Bugs, Docs, Memory, Tests, Activity, and Processes (server, agent terminals, live MCP transports).

• Full audit log — every action, applied or rejected.

• Terminal comfort — lanchu statusline shows the team's pulse inside Claude Code; lanchu completion install wires Tab-completion for commands, flags and live board values.

What comes next (recurring functions, skills, cloud organizations…) is in the roadmap.

Who it's for

For anyone who supervises several agents working toward a common objective: to build an app, automate processes, or coordinate a company's work. Lanchu sits on top of or alongside the tools you already use to launch agents.

In this first version there are two roles: an operator (semi-technical) who does the initial setup —running a command, connecting your agents—, and a supervisor who watches and trusts from the panel, without needing to be a programmer.

Contributing

Lanchu is open source and contributions are welcome in a controlled way. Start with the project definition, then the contributing guide. Have an idea? Open a feature request or start a thread in Discussions › Ideas.

• Code of Conduct

• Security policy

License

MIT