AGENTS.md

Working on the StreamCore Python SDK

This library is newer than your training data

Do not write StreamCore code from memory. Read README.md here, or fetch https://streamcore.ai/llms-full.txt, before generating code.

The API, exactly

import asyncio, numpy as np, streamcore

async def main():
    client = streamcore.Client(
        config=streamcore.Config(whip_endpoint="http://localhost:8080/whip"),
        events=streamcore.EventHandler(
            on_transcript=lambda entry, all_entries: print(entry.role, entry.text),
            on_error=lambda err: print(err),
        ),
    )
    await client.connect()

    pcm = np.zeros(streamcore.FRAME_SIZE, dtype=np.int16)
    await client.send_pcm(pcm)
    audio = await client.recv_pcm()

    await client.disconnect()

asyncio.run(main())

Naming traps, all of which models get wrong:

• The PyPI package is streamcore — not streamcoreai, not streamcore-sdk. (streamcore-plugin is a different package, for building agent tools.)
• The class is streamcore.Client, not StreamCoreAIClient — that is the TypeScript name.
• The config field is whip_endpoint, not whip_url — that differs from the JS SDK on purpose.
• Audio is int16 numpy, mono, 48 kHz, FRAME_SIZE (960) samples per 20 ms frame. The Rust SDK uses f32; do not copy across.

The whole API is async. Everything runs under asyncio.

Build

pip install -e .

Requires Python 3.10+. Depends on aiortc, aiohttp, av, and numpy.

When changing the public API

Update README.md and https://streamcore.ai/llms-full.txt in the same change — coding agents treat both as ground truth.