Open Source · 2026

Coach Claw

An open-source coaching layer for Claude Code that reads local transcripts, learns developer habits, and ships ambient tips, XP, streaks, and level-ups.

Go to site→

System

Coach Claw is a coaching layer for Claude Code. It reads local Claude Code transcripts on a schedule, learns working habits, and surfaces ambient tips, XP rewards, streaks, and level-up banners while the developer works. The daily pass is deterministic, local-only, and LLM-free — the repo contains no networking import at all. An optional weekly pass triggers Claude Code’s own insights command to refresh local sidecar files, then aggregates those deterministically; Coach itself uploads nothing.

Constraint

AI coding assistants can generate code, but they do not usually help developers improve their own operating habits over time. Useful coaching has to be specific to the developer, run quietly in the background, avoid uploading sensitive transcripts, and stay cheap enough to run every day.

Contribution

Built a local-first coaching system around 31 Python modules: a deterministic insights pass, a profile that promotes, decays, graduates, and retires tracked habits on its own, and an XP ladder of 50 levels rendered as a chess-style rating. SessionStart and UserPromptSubmit hooks inject a watch-list and pre-computed reward attribution, and Claude renders the nudge in its own voice. The engine ships twice — as an npm CLI and as a Claude Code plugin through its own marketplace — sharing one state directory with explicit coexistence handoff.

Architecture

Input

Local Claude Code transcriptsSessionStart · UserPromptSubmit hooks

System

Redaction · 20 credential shapesDeterministic pattern detectionHabit profile · XP ledger

Outcome

Watch-list injected into contextStatusline level segmentLevel-up · graduation banners

Verified

v1.0.24 on npm; Claude Code plugin at v0.1.28

MIT licensed

Zero network calls on the daily path

13 behavior patterns detected from tool-use structure alone

20 credential shapes stripped before analysis

Capabilities

Transcript analysis & detectionFilters transcripts by window, redacts credentials before a byte is parsed, and extracts 13 behavior patterns from tool-use structure alone.

Statusline renderingInstalls into an existing statusline without clobbering it, wraps a user’s own command, and renders the level segment in four variants.

Profile & progression stateConfidence scoring, tier promotion, absence-based graduation, and the XP ledger behind the level ladder.

Level naming & celebration bannersTwelve level-name themes and the banner shapes for level-ups, graduations, and regressions.

Concurrency & state filesLock-serialized read-modify-write on marker files, so concurrent sessions each consume a banner exactly once.

Dual-distribution coexistenceDetects when the npm CLI and the Claude Code plugin are both installed, hands control between them, and prunes stale caches.

Skill inventory & reward bindingInventories installed Claude Code skills, infers project scope, and binds detected patterns to scoreable actions.

ConfigurationTheme, statusline variant, and rating range, decoupled from the XP math so display changes never trigger retroactive level-ups.

Diagnostics & scheduler checksA doctor command that verifies the install, plus launchd and crontab detection across macOS and Linux.