# Cohort

A local-first, proof-carrying agent runtime that records a cryptographically verifiable replay of every model call and tool observation, letting a developer fork a past run at any turn to try a different model or prompt.

Cohort is a local agent runtime, not tied to one model vendor, that connects OpenAI-compatible or Anthropic models to an auditable, replayable, governed execution system. Its Time Machine records every request, response and tool observation from a run into a Replay Bundle carrying SHA-256 proofs; an offline Exact Replay re-verifies frame-level hashes, aggregate hashes, and request and response state transitions with no model or network calls, stopping at the first point where the recorded and actual behavior diverge, while a Fork Replay reuses recorded evidence up to a chosen turn and then switches to a live model or prompt inside an isolated Git worktree, leaving the original session untouched.

A runtime governor separates provider-reported usage from local estimates and manages context capacity, caching, cost and repeated failures; browser automation runs through a Chrome Bridge that can read the DOM, run JavaScript, click, type and screenshot; and desktop computer use is implemented through macOS Accessibility. Long-term memory must cite tool evidence and is read back on write to confirm it, and a local run log plus optional Langfuse tracing provide observability. Installed with npm.

Website: https://github.com/congchuanling-dot/Cohort