多智能体协作MCP服务器

@cyanheads/shift-mcp-server

Lightweight coordination layer for multiple AI agents working on the same codebase simultaneously. 2 Tools • 1 Resource • 1 Resource Template

Overview

Coordination layer for multiple AI agents working on the same codebase at once. Check in with a gist of your task and the files you expect to touch, see the full roster of active peers, and check out when finished.

Tools

Resources

The same roster is returned inline by shift_check_in. Modern subscriptions/listen streams receive one update per committed check-in, session update, or actual removal at the workspace-qualified URI; the bare alias is also published for the server's default workspace. Mutations from either protocol era publish through the same server event bus. This framework's legacy delivery adapter requires polling; that limitation is not inherent to the 2025 protocol. Notifications stay within one process; reads always observe committed state from other processes.

Capability reference

shift_check_in _tool

• Required gist plus optional files, workspace, and workerId (six ASCII letters/digits, case-insensitive) to update a session. IDs are stored and returned uppercase.
• Returns your session, activeWorkers, resolved workspace, and qualified statusUri. Updates preserve omitted files and the original checkedInAt; an explicit empty files array clears the declaration.
• The coordination protocol and applicable solo-worker notice are delivered once as notice in structured output and as a text trailer.
• An unknown workerId fails with reason unknown_worker (NotFound), a recovery hint, and the current roster.

shift_check_out _tool

• Required workerId (six ASCII letters/digits, case-insensitive), optional workspace and one-sentence summary.
• Returns removed, resolved workspace, qualified statusUri, and the ID/summary. An unknown or already-removed ID succeeds with removed: false and points to that roster without guessing why the ID is absent.
• Publishes a resource update only when a session actually existed and was removed.

shift://status _resource

• Returns text/markdown — the active-workers table, or "No agents are currently active." when the roster is empty
• workerId comes from shift_check_in
• The bare resource reads the process default workspace. The qualified template reads the exact identifier encoded in its URI.
• Table cells escape backslashes and pipes and flatten CRLF/LF/CR; stored values and structured output remain unchanged.

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Coordination-specific:

• Per-user SQLite roster shared across processes, isolated by workspace, durable across restarts until explicitly checked out
• The coordination protocol ships in every check-in response, so ground rules reach the agent without client-side configuration
• The active-workers table rides along with every check-in for situational awareness
• Patch semantics on session updates — only the fields provided change

Agent-friendly output:

• Both client surfaces carry the same data — structuredContent from the output schema, markdown from format()
• shift_check_in declares a typed error contract, so an unknown worker ID arrives with data.reason and a recovery hint
• Failure responses embed the current roster, so an agent recovers in the same turn instead of calling again to orient

Getting started

Add the following to your MCP client configuration file:

{
  "mcpServers": {
    "shift-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/shift-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info",
        "SHIFT_WORKSPACE": "my-project"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "shift-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/shift-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info",
        "SHIFT_WORKSPACE": "my-project"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "shift-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "MCP_TRANSPORT_TYPE=stdio",
        "-e", "SHIFT_WORKSPACE=my-project",
        "-v", "shift-state:/var/lib/shift-mcp-server",
        "ghcr.io/cyanheads/shift-mcp-server:latest"
      ]
    }
  }
}

Set the same SHIFT_WORKSPACE and use the same state directory for every agent working together. Native processes for one OS user share ~/.local/state/shift-mcp-server by default. An explicit identifier is opaque and exact-match: my-project and /path/my-project are different namespaces. Native and Docker clients can share a roster by bind-mounting the native state directory into the container instead of using a named volume.

Without SHIFT_WORKSPACE, either transport uses the real path of the enclosing Git worktree root, or the launch cwd outside Git. An empty tool workspace selector uses this default too; nonempty identifiers remain exact. Launching from a subdirectory or symlink converges on the same root. Desktop clients and install badges may launch outside your project: configure SHIFT_WORKSPACE after installation, or pass workspace in every tool call. The Desktop bundle prompts for a workspace. The Codex plugin forwards SHIFT_WORKSPACE and SHIFT_STATE_DIR from your environment; the Claude plugin exposes matching configuration options.

A shared HTTP process uses its own launch directory as the default; multiplexed clients must pass an explicit workspace for their project:

SHIFT_WORKSPACE=my-project MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.3 or higher (or Node.js v24+).
• No API keys, accounts, or external services.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/shift-mcp-server.git

2. Navigate into the directory:

cd shift-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# edit .env if you need to override a framework default

Configuration

The roster uses SQLite independently of the framework's STORAGE_PROVIDER_TYPE. State persists until check-out; restarting does not clear abandoned sessions. There is no TTL or automatic reaping. Workspace names and worker IDs are coordination selectors, not authentication boundaries.

WAL and BEGIN IMMEDIATE serialize mutations, ID allocation, and the returned roster snapshot. Startup retries SQLite lock contention within one 3-second deadline; a busy mutation also waits at most 3 seconds. Busy, corrupt, or unwritable storage fails visibly without an in-memory fallback. Use a local filesystem supporting SQLite locking for the shared state directory.

See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run:

  复制

  bun run rebuild
  bun run start:stdio
  # or
  bun run start:http
  

• Run checks and tests:

  复制

  bun run devcheck   # Lint, format, typecheck, security, packaging
  bun run test       # Vitest suites: unit, smoke, integration, fuzz
  bun run lint:mcp   # Validate MCP definitions against spec
  

Docker

docker build -t shift-mcp-server .
docker run --rm -p 127.0.0.1:3010:3010 -e SHIFT_WORKSPACE=my-project -v shift-state:/var/lib/shift-mcp-server shift-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, logs at /var/log/shift-mcp-server, and roster state at /var/lib/shift-mcp-server. Reuse the same writable volume and workspace for every coordinating container. A host bind mount must be writable by the image's bun user. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md/AGENTS.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools and resources in src/index.ts
• format() must render every field in the output schema — both client surfaces carry the same data

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.