Python bindings via PyO3 (import flowcat)

[sathish-mg]

Summary

In-process Python bindings via PyO3 — import flowcat, build a Pipeline in Python, and pass Python callables/objects as the AgentBrain (and tools) — so you can use Flowcat from Python in a single process, with no service to host.

This is on the roadmap (status: not yet started). Opening it as a tracking issue to gauge interest before prioritizing.

Today vs. this

Python can already drive Flowcat out-of-process: the RemoteBrain HTTP adapter (brain-http) runs your conversation policy as a small Python service, and the mcp client exposes Python functions as tools. That's the supported path today.

This issue is the in-process step — pip install flowcat, no HTTP hop for the control path, just Python objects.

Design intent

The binding keeps Python at turn granularity — conversation decisions and tool calls, never the per-audio-frame path — and releases the GIL during the Rust media loop, so the single-process tail-latency profile (no GC/GIL on the hot path) is preserved. The Rust media loop stays Rust; Python just decides what to say and when.

What it would unlock

• import flowcat → build a pipeline and run a call from pure Python.
• Pass a Python function/class as the brain; expose Python tools directly.
• A low-friction migration path for teams coming from Python voice stacks.

👍 Vote / weigh in

We prioritize by demand:

• React 👍 on this issue if you'd use Python bindings — that's the vote.
• Comment with what you'd build and the API shape you'd want (sync vs async, callable brain vs subclass, how tools should be registered).
• Want to help build it? Say so — it's well-scoped and high-leverage.