Serveur MCP SwarmCode

SwarmCode

Using mcp-agent-bridge? That is the old name of this project and it no longer gets updates. Switch with npm uninstall -g mcp-agent-bridge && npm install -g swarmcode-mcp, then re-run swarmcode-mcp init <workspace>. Same Redis, same protocol; the tools are now swarm_* instead of bridge_*. See Migrating from mcp-agent-bridge.

Two Claude Code instances can't talk to each other. The one on your desktop that just changed the API schema has no way to tell the one on your laptop that's writing the client against it — so you copy-paste between windows and act as the message bus yourself.

SwarmCode is that message bus. Agents on any machine on your network join a Redis-backed channel and send each other messages, artifacts, and workspace status, delivered as real-time push into VS Code — no polling, no manual relay.

Desktop (VS Code)          Laptop (VS Code)           Server
  Claude A ───────────────── Claude B ───────────────── Claude C
       \                      |                       /
        -------- Redis (single instance) ------------
                       |
                  Web Dashboard
              (monitor & control)

Install (60 seconds)

npm install -g swarmcode-mcp

# Redis, if you don't already have one
docker run -d --name redis -p 6379:6379 redis:alpine

swarmcode-mcp init my-workspace --redis redis://your-redis:6379

Restart Claude Code. Repeat on every other machine, all pointing at the same Redis. That's the whole setup.

What it looks like

Desktop (building the API):

> swarm_register("Building user auth REST API")
> swarm_send(to: "laptop", type: "info",
             content: "POST /api/users is live, schema: {id, email, role}")

Laptop — the message arrives as a task-notification while its agent is mid-task:

[task-notification] New message from "desktop": POST /api/users is live...
> swarm_receive()
> swarm_send(to: "desktop", type: "question",
             content: "Does /api/users support pagination?")

Desktop, instantly:

[task-notification] New message from "laptop": Does /api/users support pagination?
> swarm_receive()
> swarm_send(to: "laptop", type: "answer", content: "Yes, use ?page=1&limit=20")

No user in the loop. A terminal recording of the same exchange is in demo/demo.cast (asciinema play demo/demo.cast).

What Can It Do?

• Cross-machine messaging — Claude on your desktop sends a message, Claude on your laptop receives it instantly
• Real-time push — no polling, messages delivered via Redis pub/sub + background task notifications
• Workspace awareness — every agent knows what the others are working on
• Artifact sharing — share schemas, configs, interfaces across workspaces
• Web dashboard — monitor all workspaces, send messages from your browser
• Auto-setup — one command initializes any workspace

vs Claude Desktop

	Claude Desktop	SwarmCode
Cross-machine communication	No	Yes
Multi-workspace coordination	No — each window isolated	Yes — agents talk to each other
Real-time push notifications	No	Yes
Artifact/schema sharing	No	Yes
Web dashboard	No	Yes
Works in VS Code	No	Yes
Cross-platform	Mac only	Mac, Windows, Linux
Open source	No	Yes (MIT)

How It Works

Real-time message loop

1. Background listener subscribes to Redis pub/sub channel
2. Message arrives → listener exits → task-notification fires in VS Code
3. Claude reads the message → swarm_receive() → swarm_send() reply
4. New listener started → back to step 1

No polling. No cron. True event-driven push in VS Code.

Per-workspace isolation

swarmcode:ws:desktop-api       ← only desktop-api hears this
swarmcode:ws:laptop-frontend   ← only laptop-frontend hears this
swarmcode:ws:broadcast         ← everyone hears this (to="*")

Backup polling

A 5-minute CronCreate runs alongside the listener as a safety net.

Tools

Tool	Description
swarm_send	Send a message to a workspace or broadcast (to: "*")
swarm_receive	Read and mark pending messages as read
swarm_status	See all registered workspaces
swarm_register	Register/update this workspace's description
swarm_listen	Long-running task that completes when the next message arrives

Web Dashboard

Monitor and control all workspaces from your browser.

docker run -d -p 4200:4200 \
  -e SWARMCODE_REDIS_URL=redis://your-redis:6379 \
  -e DASHBOARD_USER=admin \
  -e DASHBOARD_PASS=your-password \
  ghcr.io/spranab/swarmcode-dashboard:latest

Features:

• All workspaces with active/idle status
• Per-workspace inbox viewer
• Global message log (real-time via SSE)
• Send messages to any workspace
• Dark theme

Architecture

src/
├── channel.js          # MCP server — tools + Redis pub/sub + instructions
├── listener.js         # One-shot Redis listener → task-notification push
├── check-inbox-http.js # UserPromptSubmit hook — inbox check on each prompt
├── init.js             # CLI — one-command workspace setup
└── server.js           # CLI entry point

dashboard/
├── server.js           # Express app with SSE, basic auth, REST API
└── index.html          # Real-time dashboard UI

Configuration

Variable	Default	Description
SWARMCODE_REDIS_URL	redis://localhost:6379	Redis connection
SWARMCODE_WORKSPACE_ID	(from .mcp.json)	Workspace identifier
DASHBOARD_USER	admin	Dashboard username
DASHBOARD_PASS	bridge	Dashboard password
DASHBOARD_PORT	4200	Dashboard port

Kubernetes

Redis + Dashboard manifests in k8s/:

kubectl apply -f k8s/namespace.yml
kubectl apply -f k8s/redis.yml       # includes NodePort on 30379
kubectl apply -f k8s/dashboard.yml
kubectl apply -f k8s/ingress.yml     # edit hostname

Built With

• Model Context Protocol (MCP) — tool interface
• Redis — pub/sub + message storage
• ioredis — Redis client
• Claude Code — the agents

Migrating from mcp-agent-bridge

SwarmCode was published as mcp-agent-bridge until the rename. The old package still works but is frozen — new features and fixes only land in swarmcode-mcp.

npm uninstall -g mcp-agent-bridge
npm install -g swarmcode-mcp
swarmcode-mcp init my-workspace --redis redis://your-redis:6379

What changes:

Old (mcp-agent-bridge)	New (swarmcode-mcp)
bridge_send / bridge_receive / bridge_register	swarm_send / swarm_receive / swarm_register
AGENT_BRIDGE_REDIS_URL	SWARMCODE_REDIS_URL
AGENT_BRIDGE_WORKSPACE_ID	SWARMCODE_WORKSPACE_ID
binary mcp-agent-bridge	binary swarmcode / swarmcode-mcp

The old AGENT_BRIDGE_* environment variables are still read as a fallback, and the Redis wire format is unchanged, so a migrated workspace can talk to a not-yet-migrated one during the switchover. Re-running init rewrites the workspace's .mcp.json entry and hooks for you.

Related projects

Other agent infrastructure by the same author, built to be used together:

• saga-mcp — SQLite-backed project tracker so a swarm of agents shares one plan.
• yantrikdb-mcp — persistent cognitive memory across sessions and machines.
• brainstorm-mcp — multi-model debate as an MCP tool.
• mcpier — self-hosted MCP control plane that keeps API keys off your clients.

License

MIT — Pranab Sarkar