Replaying recorded Anthropic SDK traffic offline, without touching your code

[xizhuomengcontin]

Sharing an approach for re-running code that uses this SDK without calling the API again — no changes to your source, no mocking layer.

Disclosure: I work on the tool (OrcaReplay, Apache-2.0, no paid tier).

How it works

The SDK reads ANTHROPIC_BASE_URL from the environment, so a local proxy can capture the exchange with the client left exactly as written:

client = Anthropic()        # unchanged
client.messages.create(model=..., max_tokens=..., messages=[...])

$ orca record generic-openai -- python your_script.py
info recorded run=run_… exit=0

$ orca replay last
info replaying exchanges=1 egress=blocked
info replay.done reused=1/1 exact=1 divergences=0 unmatched=0 exit=0

The replay serves the verbatim recorded bytes, so the SDK's own deserialisation and SSE handling run for real rather than being bypassed — which is where hand-written fixtures usually drift.

There is a check for this SDK in my integration matrix that records against a stub origin, kills the origin, replays, and asserts the numbers, so a break shows up as a failing build.

One thing worth knowing if you proxy Anthropic-speaking tools

Not every tool that talks to the Anthropic API reads ANTHROPIC_BASE_URL. While building harness support I found goose reads ANTHROPIC_HOST and ignores ANTHROPIC_BASE_URL entirely — a proxy addressed only by the documented variable receives nothing, and the run looks fine. Mentioning it because if you are debugging "why is my Anthropic traffic not going through my proxy", the variable name is worth checking first.

Scope

Good for: deterministic re-runs, CI without API keys or spend, chasing one bad response repeatedly.

Not: a mocking library — you record a real exchange first, you cannot author one. And egress=blocked means model-provider egress; the rest of the program still does real work. It is not a sandbox.

Happy to answer anything about the capture.