Record any coding agent. Reproduce the run byte-for-byte with the network off. Fork it from any step onto a different model and see who gets it right.

Built by the team behind OrcaRouter — one API key and one endpoint for Claude, GPT, Gemini, Grok, DeepSeek, Qwen and the rest. It is what orca setup points at by default, and what makes orca compare a single command instead of four provider accounts.

All models · OrcaCode Review · X · Hugging Face

Recording a Claude Code run, replaying it offline, then forking it onto two models

^{Real output from one session — a Claude Code run recorded, replayed with the network off, then
forked at checkpoint 4 onto two models and graded by npx tsc --noEmit. Nothing here is mocked
up.}

Install

npm i -g orcareplay
orca doctor          # checks this machine can record at all

Node 20 or newer. No native dependencies — nothing to compile, nothing to download at install time.

Three commands

orca record claude              # your agent, unmodified, doing whatever it does
orca replay last                # the same run again — no network, no tokens, no charge
orca replay last --from 4 --model claude-haiku-4-5 --ui

The third line is the one people stay for: same files, same conversation prefix, different model from step 4 onward. The model is the only variable, which is what makes the answer mean anything.

What it gives you

A timeline of what actually happened. Every model turn with its token counts and stop reason, every tool call with its arguments and result, every shell command with its exit code, every file the run touched.

$ orca show last
SEQ  KIND   WHAT                          DETAIL
12   TOOL   edit_file
14   FILE   src/auth.ts                   modified +18 −4
15   SHELL  ["npm","test","--","auth"]    /home/dev/api
16   SHELL  shell result                  exit 1 · 8.4s
17   ERROR  error

What caused what. Every edge says whether the recorder watched it happen or orca derived it just now, and names the rule either way.

$ orca graph last --to 17
FROM              TO                KIND      WHY
12 tool.call      15 shell.exec     recorded  causes
15 shell.exec     16 shell.result   recorded  shell result answers its exec
16 shell.result   17 error          recorded  causes

The same task on several models, graded by a command you choose. Not a model marking its own homework — the verdict is the exit code of whatever you tell it to run, so the same prompt costs you a number you can act on rather than an opinion.

$ orca compare last --from 4 --models claude-sonnet-5,claude-haiku-4-5 --verify "npx tsc --noEmit"
MODEL                      VERDICT  TOKENS   COST       WALL
claude-sonnet-5            pass     124/429  $0.006807  12.2s
claude-haiku-4-5-20251001  pass     184/650  $0.003434  14.3s

Both passed here; the interesting column is what each one cost to get there.

Plus a browser timeline (orca ui), a single self-contained HTML file you can attach to an issue (orca export), and shareable cards of one causal chain (orca export --card).

Which agents

Claude Code, Codex, OpenCode, grok-cli, the Anthropic and OpenAI Agents SDKs, the Vercel AI SDK, and any harness that reads a base-URL environment variable — including ones this project has never heard of, via ORCA_BASE_URL_VARS. Harnesses that refuse to be redirected can be captured through TLS interception with a certificate authority minted for that one run and thrown away after it.

orca doctor reports which of them it can find on your machine.

Full documentation

github.com/Continuum-AI-Corp/OrcaReplay — the worked bug hunt, the trace format spec, the programmatic API, CI usage, what is stored and where, and the privacy model.

Apache-2.0. The trace format spec is CC BY 4.0, so anything can read or write these traces.