Record any AI agent CLI's loop with one env var. Replay it offline — key revoked, $0.
loop-replay is a local reverse proxy that records an agent's full loop at the provider API boundary — every model request and response, including tool_use / tool_result turns — and replays it deterministically. No SDK. No code changes. Point ANTHROPIC_BASE_URL at it and run your agent.
- No integration seam. Works with agent CLIs you don't control (Claude Code, Codex CLI, aider, opencode) — one env var, zero instrumentation.
- $0 deterministic replays. Re-run your agent's loop in CI with the API key revoked. Harness-code changes get regression-tested against the tape.
- Golden tasks, outcome graders. Pin real tasks, grade final workspace state (never text snapshots), and diff harness config A/B with a tape-backed control leg.
Pre-alpha — under active construction. Nothing to install yet. First public demo target: record a real Claude Code session, revoke the key, replay the identical transcript offline.
MIT