Replaying a CAMEL ChatAgent run offline — plus a packaging note on camel-ai 0.2.90 and mcp #4326
xizhuomengcontin
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Verified against camel-ai 0.2.90 today. Two things: a packaging problem I hit on the way in, and the thing I was actually trying to do.
Disclosure: I work on the tool (OrcaReplay, Apache-2.0).
First, the packaging note — this is the useful half
A clean
pip install camel-aiinto a fresh Python 3.12 venv fails on any import ofcamel.agents:FastMCPlives atmcp.server.fastmcpin currentmcpreleases, so whatever version pip resolves no longer satisfies that import. Pinningmcp==1.9.4fixes it:pip install camel-ai "mcp==1.9.4"Nothing to do with my tool — it happens on a bare install — but it means a new user's first
from camel.agents import ChatAgentfails. Worth an upper bound onmcpin the dependency spec, or moving the import.Then the thing itself
A real
ChatAgentoverOPENAI_COMPATIBLE_MODEL, recorded and replayed with the model server gone:Nothing is installed into CAMEL — the origin comes from the environment, so
ModelFactory.create(...)is unchanged. Useful for re-running an agent while you change the system message or the toolkit, for CI, and for chasing one bad trajectory without paying for it each time.(Unrelated but noted while running: CAMEL warns
Unknown model 'qwen3.5:0.8b': context window size not defined. Defaulting to 999_999_999for custom endpoints. Sensible fallback, just surprising to read.)Scope
egress=blockedon replay means model-provider egress; recorded tool calls still execute for real, so it is not a sandbox. And this captures the HTTP going to the model, not CAMEL's internals — not a substitute for step-level tracing.Happy to open a separate issue for the
mcppin if that is more useful than a note here.All reactions