How it works
OpenPlan exposes three MCP tools. Each does exactly one thing. Together they form a complete project lifecycle.
01
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.
plan( goal="Build a landing page", context="Astro + Tailwind", project="marketing-site" )
02
Record phase completion with actual cost. Returns deviation, hazards, and next pending phase. Correct previous costs with correct=. Check status anytime with no args.
checkpoint( phase="Scaffold", actual_cost=2100, route_id="uuid" )
Correct a phase cost:
checkpoint( phase="Scaffold", correct=1950, route_id="uuid" )
03
Full retrospective with deviation analysis, accuracy metrics, cost/path learning, and diagnostics. Calibration data feeds back to the global Mesh.
review( route_id="uuid" )
Why OpenPlan
Every principle serves one goal: make agents better at estimating, tracking, and learning from software projects.
Agent-native interface
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
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
Data lives in a local SQLite file. Created on first call. Zero configuration. Works offline. No cloud dependency.
Structured errors, not crashes
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
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
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
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
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
Every agent contributes. Every agent benefits. The more projects you ship, the sharper the estimates get.
PRICING
The server is free and always will be. Pro gives you a personal view into the global calibration data — estimates tuned to how you actually build.
One command. Zero setup. Works with every MCP host.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。