OpenPlan — Waze for AI agents MCP SERVER · v0.1.22<br>Waze for AI agents.<br>Plan, track, and learn from software projects.
$ npx @openplan/mcp Install now ↗ GitHub ↗
How it works<br>Three calls. One cycle.
OpenPlan exposes three MCP tools. Each does exactly one thing.<br>Together they form a complete project lifecycle.
01<br>plan()<br>Decompose a goal into costed phases with estimates from the global calibration pool. Returns a route with phases, confidence intervals, evidence (hazards), personal bias, and archived routes.<br>plan(<br>goal="Build a landing page",<br>context="Astro + Tailwind",<br>project="marketing-site"
02<br>checkpoint()<br>Record phase completion with actual cost. Returns deviation, hazards, and next pending phase. Correct previous costs with correct=. Check status anytime with no args.<br>checkpoint(<br>phase="Scaffold",<br>actual_cost=2100,<br>route_id="uuid"<br>) Correct a phase cost:<br>checkpoint(<br>phase="Scaffold",<br>correct=1950,<br>route_id="uuid"
03<br>review()<br>Full retrospective with deviation analysis, accuracy metrics, cost/path learning, and diagnostics. Calibration data feeds back to the global Mesh.<br>review(<br>route_id="uuid"
Why OpenPlan<br>Designed for the way agents work.
Every principle serves one goal: make agents better at estimating,<br>tracking, and learning from software projects.
Agent-native interface<br>Three MCP tools — plan, checkpoint, review. No REST API to learn, no SDK to import, no auth to set up. Every tool returns structured JSON the agent can act on.
One call per action<br>Plan once. Checkpoint after each phase. Review at the end. Minimal round trips, maximum signal. Every call has one purpose and one response.
Local-first, no setup<br>Data lives in a local SQLite file. Created on first call. Zero configuration. Works offline. No cloud dependency.
Structured errors, not crashes<br>Every response is parseable JSON. Invalid arguments, missing routes, rate limits — all return typed error codes. The agent never sees a stack trace.
Learning that compounds<br>Every checkpoint trains the model. Accuracy, deviation, and personal bias are tracked per action. SQL aggregates replace ML pipelines — no vectors, no training data, no black boxes.
Global mesh, anonymous<br>Calibration events sync to a shared pool with built-in poisoning defenses (MAD filter, Bayesian shrinkage, per-key rate limiting). Every agent benefits from every other agent's experience.
Graceful by design<br>Cost probes return null when unavailable. Sync failures log silently. Missing routes return NOT_FOUND, not an exception. The agent's flow never breaks.
Agent loop<br>Plan. Execute. Checkpoint. Review.
One call to plan. One call per phase to checkpoint. One call to review. That's the whole loop — and every loop makes the next one more accurate.
plan() Decompose goal into costed phases checkpoint() Record cost & deviation checkpoint() — status Full route state, no phase needed review() Retrospective & diagnostics
Architecture<br>Local server. Global mesh.
Every agent contributes. Every agent benefits. The more projects you ship, the sharper the estimates get.
PRICING<br>Free for everyone. Better with Pro.
The server is free and always will be. Pro gives you a personal view<br>into the global calibration data — estimates tuned to how you actually build.
Free $0<br>3 MCP tools: plan, checkpoint, review<br>Local SQLite (unlimited, offline)<br>Global calibration pool (100 pulls/day)<br>Pool-only baselines — same estimates for everyone<br>Mesh sync toggle via CLI<br>Anonymous — no account needed<br>Install now ↗<br>Pro $9 /month<br>Everything in Free<br>Personalized estimates — learns from your actual history<br>Unlimited baseline pulls, priority queue<br>Export your data (JSON, CSV, Markdown)<br>Roaming — authenticate once, Pro follows you<br>$ openplan subscribe
Built for the agent loop.<br>One command. Zero setup. Works with every MCP host.<br>Install now ↗ View on GitHub ↗<br>$ npx @openplan/mcp