OrcaReplay

OrcaReplay records a coding agent's run so it can be replayed exactly, offline, and forked onto a different model. It targets agent debugging: instead of reading cost and token dashboards, you get the run back as a file and can step through which model turn or tool call caused a problem.

It works without modifying the agent. Because model APIs are stateless and an agent resends the whole conversation each turn, a local proxy placed in front of the model sees every request, streamed response, tool call, and tool result. Further capture layers record shell exit codes and timing through a PATH shim, MCP traffic through a JSON-RPC tee, per-turn workspace snapshots in a shadow git index, a fetch hook for agents with a hardcoded origin, and an opt-in TLS intercept for agents that cannot be redirected. Everything lands in one timeline in an open trace format.

Exact replay, fork, and compare use the same proxy with a cursor: up to the cursor responses come from the recording, and after it they come from a live model. Exact replay restores the recorded filesystem and calls no model. A fork continues from a chosen checkpoint on another model, and compare forks one run onto several models and grades each with a command such as a test suite, so the model is the only variable.

Supported harnesses include Claude Code, Codex CLI, the OpenAI Agents SDK, the AI SDK, OpenClaw, OpenCode, Goose, LangGraph and LangChain, Haystack, OpenHands, CrewAI, Aider, Browser Use, and Hermes, plus any tool that reads a base-URL variable. Runs can be pushed to and pulled from an OrcaRouter gateway, and every command emits JSON so agents and CI can query runs, with an MCP server that lets an agent read and replay its own runs. The project describes itself as early.