🐝 Swarm Protocol

Coordination protocol for agent-first teams. No UI. No sprints. No Jira. Just state sync.

Status: Alpha — building in public. 19 MCP tools implemented, integration tests passing.

The Coordination Loop

Drop COORDINATION.md into your repo's CLAUDE.md and agents coordinate automatically:

Agent starts session
  → get_team_status        "What's in flight?"
  → claim_work             "I'm taking this — here are my files"
  → check_conflicts        "Anyone else touching src/api/router.ts?"
  → heartbeat              "Still working" (every 10-15 min)
  → complete_claim          "Done — this unblocks intent_xyz"
                              ↳ intent_xyz status: blocked → open
                              ↳ Next agent picks it up automatically

Zero Slack messages. Zero status meetings. Agents coordinate through shared state.

What This Is

A headless coordination layer exposed as an MCP server — the same protocol Claude Code already speaks. No REST API. No dashboard. Agents query it to see what's in flight, claim work, detect file conflicts, and hand off unblocked tasks.

This is a coordination protocol, not a project management tool. Jira was built for humans clicking buttons. This is state synchronization infrastructure for teams where agents are the primary development interface.

The CLAUDE.md integration pattern is as important as the server itself. Agents coordinate automatically without humans configuring anything.

How This Relates to Agent Teams, Code Review, and Other Multi-Agent Tools

These tools are complementary, not competing. They operate at different layers:

• Claude Code Agent Teams coordinates agents within a single session — one lead, parallel teammates, shared task list. Intra-session.
• Claude Code Review dispatches parallel review agents against a single PR. Intra-PR.
• VS Code multi-agent runs Claude, Codex, and Copilot agents side by side for one developer. Intra-machine.
• AgentConductor dynamically adjusts agent communication graphs per coding problem. Intra-problem.

Swarm Protocol coordinates across sessions — multiple humans, each running their own agents, on a shared codebase. Inter-session, inter-human.

An agent uses Agent Teams internally for its own subtask parallelism. It uses Swarm Protocol's MCP tools to know which work to pick up, which files to avoid, and what context the previous agent left behind.

The single-player problem is getting solved. The multiplayer problem hasn't started.

Who Is This For

Teams of 2+ developers where AI agents (Claude Code, etc.) are the primary development interface. If your workflow is "open terminal → tell Claude Code what to build → review the PR" and teammates are doing the same thing at the same time — this is the missing layer.

Not for: solo devs running multiple agents in parallel (tools like CCPM and 1Code solve that). Not a Jira replacement. Not a project management tool.

The Problem

Every existing tool solves: "I'm one developer running 3-5 agents in parallel, how do I prevent them from colliding?"

Nobody is solving: "We're a team of 8 humans, each working through agents, across the same codebase. How do we know what's in flight, avoid stepping on each other, and automatically hand off unblocked work?"

That's single-player vs. multiplayer. Different product category. And it's not a project management problem — it's a state synchronization problem.

What happens without coordination:

• Two agents edit the same files → ugly merge conflicts
• Completed work sits idle — no one picks up the next dependent task
• Context is lost between agent sessions
• No shared state of what's done, in progress, or blocked

See LANDSCAPE.md for the full competitive breakdown.

Four Primitives

Architecture

Stack: Node.js + TypeScript · PostgreSQL (raw SQL, no ORM) · @modelcontextprotocol/sdk

Design decisions:

• Conflicts are advisory, not enforced — no file-level locking
• Auth is trust-based in v1 (claimed_by is a string the agent passes)
• Single PostgreSQL instance (designed for <1000 users)
• Polling via MCP tools, no WebSocket subscriptions
• Protocol over product — minimal and composable by design

Key Tools

check_conflicts — Prevent file collisions before they happen

→ check_conflicts({ files: ["src/api/router.ts", "src/middleware/auth.ts"] })

← { conflicts: [{
     file: "src/api/router.ts",
     claimed_by: "anna",
     intent: "Refactor API error handling",
     claim_id: "claim_abc123"
   }]}

get_context — One call, full onboarding

→ get_context({ intent_id: "intent_abc123" })

← { intent: { title, description, constraints, acceptance_criteria },
    parent: null,
    dependencies: [{ id: "intent_xyz", status: "done" }],
    active_claims: [{ file: "src/middleware/rateLimit.ts", claimed_by: "pawel" }],
    recent_signals: [...],
    team_conventions: "Use Prettier. Tests required. No console.log in production." }

complete_claim — Done. Dependents auto-unblock.

→ complete_claim({ claim_id: "claim_def456", unblocks: ["intent_xyz"] })

← intent_abc123: claimed → done
   intent_xyz: blocked → open  (all dependencies met)
   signal: completion created

All 19 Tools

See SPEC.md for full parameter specs and data model.

Quick Start

Prerequisites: Docker, Node.js 22+

# Clone and set up
git clone https://github.com/phuryn/swarm-protocol.git
cd swarm-protocol

# Start PostgreSQL
docker compose up -d

# Build and test
npm install
npm run build
npm test

Add to Claude Code (~/.claude/config.json):

{
  "mcpServers": {
    "swarm-protocol": {
      "command": "node",
      "args": ["/path/to/swarm-protocol/dist/index.js"],
      "env": {
        "DATABASE_URL": "postgresql://postgres:postgres@localhost:5432/swarm_protocol"
      }
    }
  }
}

Enable automatic coordination — copy claude-md/COORDINATION.md into your repo's CLAUDE.md. That's it.

What's Not in v1

Intentional omissions, not missing features:

• No web UI — agents and terminal are the interface
• No auth — trust-based identity. Auth layer comes when it needs to.
• No real-time subscriptions — MCP polling is sufficient
• No notifications — Slack/Discord integration is a natural v2 extension
• No file locking — conflicts are advisory by design
• No multi-repo — single repo per team assumed

Looking for Co-Builders

The market for agent-team coordination is tiny today and enormous in 12-18 months. Every team adopting Claude Code, Codex, or similar tools will hit this exact problem the moment they scale past one developer.

This isn't a side project looking for drive-by PRs. It's a category that needs to be built. I'm looking for people who want to own parts of this protocol — lead feature areas, review PRs, shape the design.

Where to plug in:

• Tool groups (src/tools/) are natural ownership boundaries — each one is a self-contained module with its own tests
• The protocol design itself — the four primitives are v1. What's missing? What's wrong? Open an issue and make the case.
• Adapters — SQLite backend, auth layer, Slack signal forwarding, dashboard read-model. Each is a standalone contribution.
• The CLAUDE.md pattern — better coordination instructions, support for other agents beyond Claude Code

The raw SQL + no-framework design is intentional — fork it, swap PostgreSQL for SQLite, add auth, build custom tools.

See CONTRIBUTING.md for guidelines. Or just open an issue with your idea.

Docs

License

MIT