OrcaReplay
A local AI agent debugger records coding runs, replays saved model responses without model calls, and compares models from the same checkpoint.
OrcaReplay is a local AI agent debugger for developers investigating failed coding runs or testing how another model handles the same task. It captures what happened beyond the chat transcript, including failed commands and changed files.
- Run inspection: Review model conversations, tool results, shell exit codes, timing and file changes in a timeline ordered by when events happened.
- Recorded replay: Repeat a run using saved model responses without calling a model; tool calls still execute, so replay isn't a sandbox and tools can still access the internet.
- Model comparisons: Continue from a checkpoint with the recorded conversation and workspace state, switch models, and assess the resulting runs with project checks.
- Agent capture: Record Claude Code without modifying its code, or use opt-in interception to capture Codex CLI sessions signed in through a ChatGPT subscription.
- Cause inspection: View connections between model responses, tool calls and file changes, with recorded relationships distinguished from inferred ones.
OrcaReplay is open source under Apache 2.0. Recordings live in files on your machine; live recording and model forks send requests to a model provider. OrcaRouter is the default cloud gateway used by setup and provides access to Claude, GPT, Gemini, Grok, DeepSeek and Qwen. Recorded replay avoids model requests, but it doesn't block every network request made by the agent's tools.
