OrcaReplay
A time-travel debugging tool for AI coding agents that records, replays offline, and forks agent runs to compare different LLMs on the exact same task.
🔗 Visit OrcaReplay
Description
When an AI coding agent does something wrong, it's hard to know why — the agent's "thinking" happened once, live, and by the time you notice the bug that exact moment is gone. OrcaReplay solves this by recording everything the agent did like a flight recorder, so you can rewind and watch the run again offline for free, or even swap in a different AI model at any checkpoint to see if it would have handled the same situation better.
OrcaReplay is an open-source (Apache-2.0) Node.js tool that captures a coding agent's full execution — proxy traffic, shell commands, MCP tool calls, filesystem changes — and lets you replay it byte-for-byte without hitting the network or paying for tokens again. It supports capture from Claude Code, Codex CLI, the OpenAI Agents SDK, Vercel AI SDK, grok-cli, LangGraph/LangChain and custom agents, and its standout feature is "model forking": replay up to a checkpoint, then continue with a different LLM to compare behavior on an identical prior context. It also builds causal graphs showing what caused what during a run, exportable as single-file HTML, SVG or PNG.
💬 Our review
The short version: OrcaReplay is the first tool that treats "did switching models actually help" as an answerable question instead of a vibe, by letting you fork a recorded run and swap the LLM at any checkpoint.
Regular agent observability platforms like Langfuse or LangSmith show you cost and latency after the fact, but they can't rerun the exact same agent state with a different model — OrcaReplay's checkpoint-forking is the differentiator. The catch: only Claude Code has been validated end-to-end; support for the other agents relies on adapter fixtures, so treat those as beta. It's free and open-source, worth adopting the moment you're debugging a flaky agent run or trying to justify a model switch with evidence instead of anecdotes; skip it if you just want basic cost/latency dashboards, where a lighter observability tool is simpler to set up.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Gratuit, licence Apache-2.0
Pros
Rejeu byte-for-byte hors ligne sans coût de tokens
Fork de modèle depuis un checkpoint pour comparer des LLM sur un contexte identique
Graphe causal visualisant les dépendances d'une exécution
Cons
Seul Claude Code est validé de bout en bout, les autres reposent sur des fixtures adaptateur
Sessions tapées à la main non reproductibles byte-for-byte
Export PNG/GIF nécessite Chromium en option