I'll take another look but from what I perceived all attempts to start a thread included tool calling in the payload.

I couldn't seem to get any message through without tool calling instructions in the payload. What you're describing sounds exactly like what I attempted.

I tried something like over 6 different variations of model configs with restarts of Zed in-between. The documentation and what Zed tries to configure are different as well. The fields don't match up with the built in type checking. I tried "openai" with the endpoint configured, "openai_compatible", and even "openrouter" hoping the REST signatures would be match well enough. Each configured with various fields to turn tool calling off and every single request that hit the REST server had tool calling.

You should have a drop-down for selecting a tool profile next to the drop-down for model selection. Select "Minimal" for no tools.

"New text thread" should also have no tools I believe.

https://zed.dev/docs/ai/agent-panel#profiles

Text threads include no tooling. They are plain markdown texts you can freely edit (including llm's responses) so you can fix small stuff yourself without annoying, context-polluting back and forths with the llm, at the potential cost of possible caching optimisations.