kaizen

Kaizen watches your coding agents work, locally, across Cursor, Claude Code, Codex, OpenClaw, Goose, OpenCode, and Copilot.

At the end of the week, it tells you what's wasting tokens, causing loops, or making agents worse in this repo.

Then it lets you test the fix: "add this rule", "split this file", or "delete this unused skill" becomes measurable instead of vibes.

Demo

https://github.com/user-attachments/assets/64e2739a-6102-4abb-9ce7-9d4046dd56ca

Install

You need Rust 1.95+ from rustup.

cargo install kaizen-cli --locked

The command installs the kaizen binary into ~/.cargo/bin or $CARGO_HOME/bin. If your shell cannot find it, add that directory to PATH and open a new terminal.

Developing Kaizen itself? Use ./scripts/install-local.sh from a clone, or cargo install --path . --locked. Detailed install notes live in docs/install.md.

Quick Start

cargo install kaizen-cli --locked
cd my-repo
kaizen init
# use your coding agent for a day...
kaizen summary
kaizen retro

kaizen init creates local storage under .kaizen/ and wires supported agent hooks idempotently. Re-running it is safe; originals back up under .kaizen/backup/.

How the Loop Works

Observe. Kaizen tails agent transcripts and hook events into workspace-local SQLite. It can watch Cursor, Claude Code, Codex, OpenClaw, Goose, OpenCode, and Copilot without running the model itself.

Summarise. kaizen summary and kaizen metrics roll sessions up by agent, model, cost, tool use, and repo facts. The metrics pass also indexes file-level and graph facts so retros can talk about this codebase, not only token totals.

Propose. kaizen retro runs deterministic heuristics and groups bets by confidence and action type: one high-confidence bet, up to two investigations, and up to two quick hygiene fixes.

Measure. kaizen exp new can bind a change to git history and compare control vs treatment sessions with bootstrap confidence intervals.

Nothing leaves disk unless you configure sync or a provider query. Team paths redact secrets, env vars, absolute paths, and git emails before upload.

Why Kaizen

Kaizen is not a dashboard. It is an opinionated feedback loop: capture → summarise → propose change → measure.

Advanced Features

• HTTP proxy: kaizen proxy run logs Anthropic API calls with precise token counts and optional context policy trimming.
• Provider queries: PostHog and Datadog rollups can be read with --source provider or merged with local rows via --source mixed when [telemetry.query] is configured.
• Redacted sync: shared endpoints receive batches only after local redaction and UUIDv7 deduplication.
• A/B experiments: kaizen exp new --bind git classifies later sessions by commit boundary and reports treatment deltas.
• MCP tools: most read/report commands are exposed to agents; shell-only commands stay in the CLI.
• Local depth: optional test/lint snapshots and CPU/RSS samples improve retro signals without shipping raw command output off disk.

Docs

Status

Pre-1.0: breaking changes may appear in minor versions. Follow CHANGELOG.md for release notes.

License

AGPL-3.0-or-later. Contributions are licensed under the same terms. Security disclosures: SECURITY.md.