ContextPool

The problem ContextPool targets is real and under-served: AI coding agents have no durable memory of your project. Cursor, Claude Code, Windsurf, and Kiro each scope context to the session, so knowledge you build up — why you picked libsql over rusqlite, which async pattern triggered a tokio panic, that macOS keychain calls block in an MCP subprocess — evaporates when the chat ends. ContextPool intercepts that loss in three steps: install a static binary via curl, run cxp init so terminal replay scans past Cursor and Claude Code sessions and an LLM extracts insights, then let the MCP server serve relevant past context at session start. What it gets right is the shape of the memory. It stores actionable engineering knowledge — categorized as bug, fix, decision, and gotcha — rather than conversation summaries, which is the difference between a useful recall and noise. The four example extractions on the homepage are exactly the kind of project-specific detail a fresh agent cannot reconstruct from your code alone. The operational design is equally thoughtful. Claude Code is zero-config and reuses your existing Claude Code auth, so there is no API key to create. Other editors need one JSON entry, which is a low bar. Project IDs come from your git remote URL, so every teammate on the repo resolves to the same memory namespace automatically. Multi-backend LLM routing runs a fallback chain — Claude CLI → Anthropic API → OpenAI → NVIDIA — so extraction keeps working if one provider is unavailable, though that chain may add latency during extraction. API keys go in the system keychain with a safe file fallback. On the privacy side, transcripts stay local, secrets are stripped before LLM processing and again before sync, and only insights — never full transcripts — are synced when you opt in. Where it does not fit: there is no GUI, so if you want a dashboard to browse and curate stored context, this is not that tool. It is CLI-only. A brand-new project has no session history to mine, so day-one value is near zero — the tool compounds with use. It is explicitly for people who already run an AI coding agent; non-developers get nothing from it, and single-session disposable projects will never generate enough history to matter. Teams that require a fully managed cloud-only service should look elsewhere, since the local-first model is the whole point here. And if you are happy maintaining a hand-written context file you paste at the top of each session, you have a working alternative — ContextPool just automates the capture and recall you are doing manually.