Hi Product Hunt! OpenPlan gives AI agents a memory for costs. Three MCP tools — plan, checkpoint, review — that turn every project into a learning loop. I noticed a pattern: AI coding assistants ship code fast but can't estimate to save their life. Every project starts with "3 days" and ends with "well, that was quick." OpenPlan fixes that by tracking actual costs and feeding them back into future estimates. The more you use it, the sharper the predictions get. The server is MIT, runs locally via stdio, and works with any MCP host. Cost probes are community-driven and you are welcome to contribute. Adapters for OpenCode, Claude, Cursor, Codex, or any CLI command. If your host exposes cost data, the adapter pattern is simple. The calibration pool is fully anonymized. Tokenized labels, random UUIDs, no source code or project names ever leave your machine. Free tier includes local SQLite, mesh sync, and pool-based estimates. Pro adds personalized estimates, unlimited pulls, and data export. Would love feedback — especially if you've tried other approaches to estimation in AI-assisted dev.

OpenPlan

Waze for AI agents. Plan, track, and learn from projects.

Waze for AI agents. Plan, track, and learn from projects.