The agentic task substrate for AI coding agents. An MCP server that turns GitHub Projects v2 into a fully autonomous project management platform — AI agents self-assign work, track progress, review each other, and ship code, all backed by GitHub-native storage.
Use case: You have AI agents (Claude Code, Codex, Cursor, Windsurf, Roo). They need a task backbone. This is it.
Overview
This MCP server implements the Model Context Protocol to provide a complete agentic project management layer over GitHub Projects v2. AI agents register, self-assign tasks, coordinate via heartbeats, submit work products for review, and operate within token budgets — all through 16 compound tools exposing 152 actions. Human project managers get AI-powered PRD generation, sprint planning, issue triage, and roadmap creation. Everything is backed by GitHub-native storage (issues, project fields, comments) — no external infrastructure required.
Why This Exists
AI coding agents are powerful but stateless — they don't know what to work on next, can't coordinate with other agents, and have no persistent task memory. This MCP server solves that by turning GitHub Projects into an agentic operating layer:
| Problem | Solution |
|---|---|
| Agents don't know what to work on | Task checkout with priority/skills/deadline strategies |
| Multiple agents step on each other | Atomic claims with TOCTOU guards + review workflow |
| No visibility into agent work | Heartbeat monitoring + activity dashboard + work product tracking |
| Agents burn unlimited tokens | Per-agent token budgets with warning thresholds and hard stops |
| Ideas don't become tasks | PRD→Tasks→Issues pipeline with AI-powered breakdown |
| Sprint planning is manual | AI sprint planning with capacity analysis and risk assessment |
| Issues pile up untriaged | AI-powered triage with auto-labeling and priority assignment |
| Need external databases | Zero infra — all state in GitHub issues, projects, and comments |
Works With
| Agent / Client | Integration |
|---|---|
| Claude Desktop | Native MCP client — add to claude_desktop_config.json |
| Claude Code | MCP server via stdio transport |
| Codex | MCP-compatible — register as tool provider |
| Cursor | MCP server integration |
| Windsurf | MCP server integration |
| Roo Code | MCP server integration |
| VS Code + Copilot | Via MCP extension |
| Custom harnesses | Any MCP-compatible client — see examples/basic/agent-loop.ts |
What Makes This Special
- Agentic-First Design: Built as a task substrate for AI agents, not a human-facing PM tool retrofitted with an API
- 16 Compound Tools (152 actions): Progressive-disclosure API — agents see 16 tools, drill into 152 actions via
discover_tools - Multi-Agent Swarm: Agents register, claim tasks atomically, heartbeat, review each other, and converge projects
- PRD→Production Pipeline: Idea → PRD → tasks → GitHub issues → milestones → sprints — fully automated
- Zero External Infrastructure: All state lives in GitHub (issues, projects, comments) — no database, no Redis, no S3
- AI-Powered Everything: Triage, sprint planning, roadmap generation, complexity analysis, duplicate detection — all AI-augmented
- Token Budget Enforcement: Per-agent spending limits prevent runaway AI costs
- Self-Healing: Auto-reclaim tasks from crashed agents, stale heartbeat detection, registry cleanup
No comments yet.