openplan-mcp 0.8.3

Waze for AI agents planning — MCP server that improves cost estimates by learning from every project

OpenPlan

Waze for AI agents planning — an MCP server that helps AI agents plan software projects efficiently by learning from every agent's cost data.

How it works

AI agents use OpenPlan's tools to track project phases, costs, and outcomes. Every start() and complete() call generates calibration data that improves estimates for every other agent — like Waze uses every driver's trip data to give better ETAs.

start → complete × N → verify → recommend

Quick Start

Via PyPI

pip install openplan-mcp

Then add to your MCP host config:

{
  "mcp": {
    "openplan": {
      "type": "local",
      "command": ["uvx", "openplan-mcp"]
    }
  }
}

Or use uvx directly (no install needed):

{
  "command": ["uvx", "openplan-mcp"]
}

CLI

openplan                  # Start MCP server
openplan auth login       # Authenticate with GitHub for Pro tier
openplan auth logout      # Remove credentials
openplan auth status      # Show authentication state
openplan subscribe        # Start Pro subscription ($10/mo)
openplan status           # Show OpenPlan status

Tools

Tool	Description
start	One-call project kickoff: parses goal into phases, estimates costs from global baselines
complete	Mark a phase done, attaches evidence, auto-traverses to next phase
act	Traverse, branch, verify, set status, abandon, prune, revert
recommend	Best next step with A* path, project health, cost estimates
export	Export full graph as JSON / GraphML / matrix

Architecture

The MCP server runs locally. Calibration data syncs to api.openplan.cc (optional, anonymous by default). The cloud aggregates anonymized cost data across all users — every project improves estimates for everyone.

  MCP host (OpenCode / Claude Desktop / Cursor)
       │
  openplan MCP server (stdlib, uvx openplan-mcp)
       │
       ├── local SQLite (your projects, always works offline)
       │
       └── api.openplan.cc (global calibration pool, optional)

Data Privacy

Only {project_type, action, expected_cost, actual_cost, outcome} is shared — no source code, no project names, no file paths. Anonymous by default. GitHub OAuth for Pro features.

License

MIT