agent-coordination-mcp

An experimental local MCP server for coordinating installed CLI agents across projects that use file-based task boards, locks, and shared status files.

The goal is not to replace the ai-agent-teamwork workflow. It is to expose that workflow as a small MCP control plane so a primary agent can see which CLI agents are available, assign work, inspect project coordination state, and keep durable assignment records.

Why This Works

The workflow works if the MCP server stays narrow:

• MCP is the coordination surface, not the editor.

• Project state remains in plain files owned by each project.

• CLI agents continue to run as separate tools with their own approval and sandbox behavior.

• Assignment and heartbeat records are explicit JSON state, not inferred from hidden sessions.

The risky part is supervising long-running CLI processes. This first slice records assignments and generates dispatch intent; later slices can add process launch adapters per CLI once the approval and lifecycle model is clear.

Related MCP server: kanban-lite

Current Tools

• list_cli_agents - detect known local CLI agents on PATH.

• get_project_status - summarize .agent-tasks.json, .agent-manifest.json, and .agent-status.md for a project.

• list_assignments - read active and historical assignment records.

• assign_task - record a task assignment to a detected CLI agent.

• update_assignment_status - update assignment status and notes.

Install

uv sync
uv run agent-coordination-mcp

Dynamic MCP Proxy Entry

Add this to /home/stephen/dynamic-mcp-proxy-server/user.catalogue.json:

{
  "name": "agent-coordination",
  "description": "Local MCP control plane for detecting CLI agents and coordinating file-based project task boards",
  "command": "uv --project /home/stephen/projects/agent-coordination-mcp run agent-coordination-mcp",
  "tags": ["agents", "coordination", "mcp", "cli", "local"],
  "tech_stack": ["python", "mcp", "cli-agents"],
  "runtime": "stdio",
  "env_vars": []
}

Planned Slices

1. Inventory and assignment tracking.

2. ai-agent-teamwork task board adapters.

3. CLI-specific dispatch adapters for opencode, codex, gemini, claude, and other installed agents.

4. Process/session tracking where supported by the CLI.

5. Dynamic proxy integration and research tools such as devto-mcp-server.

See docs/ROADMAP.md for the current implementation plan. The next slice is capability-aware inventory before automated dispatch.

Non-Goals

• No hidden project edits by the MCP server.

• No generic shell execution tool.

• No automatic force-unlocking or stale-task rewrites without explicit tool calls.

• No assumption that every CLI supports resumable sessions.