A self-hosted voice agent runtime. Speech-to-text, the LLM and text-to-speech run together on one machine and stream into each other, so a reply starts playing while it's still being generated.

On an RTX 3090 with a 7B model: about 490 ms of processing once a turn ends, or 991 ms stopwatched from your last syllable — the difference is a silence wait you can configure. 127 tokens/sec, interruptions honoured mid-sentence.

Quickstart

Requires Python 3.11–3.13.

pip install fusion-runtime

An agent is one file. This is the whole thing:

# agent.py
from fusion_runtime import Agent, LLM, STT, TTS, Turns

agent = Agent(
    name="shopkart-orders",
    prompt="You are the order line for ShopKart. Keep answers to one short sentence.",
    stt=STT("whisper-tiny.en"),          # or "whisper-small" for better accuracy
    llm=LLM("qwen2.5-0.5b-q4", max_tokens=256),
    tts=TTS("kokoro-v1.0", voice="af_heart"),
    turns=Turns(wait_ms=500, interrupt_after_ms=300),
)

frun models pull agent.py     # exactly the models it names, nothing else
frun up agent.py              # add --reload to restart on every edit

Then talk to it from a second terminal:

pip install "fusion-runtime[talk]"
frun talk

That is the whole loop — one file, two commands, a conversation. Talk over the agent to interrupt it.

frun up with no file runs a default agent if you just want to hear it work, and frun doctor checks libraries, GPU, models and audio and says how to fix what it finds.

In a browser instead

The runtime serves a browser client at http://localhost:8000 — the same one you would embed in your own page.

With no keys configured, open it and click Talk. With keys configured (FUSION_ACCEPTED_KEYS), a page can't hold a secret, so it needs a short-lived session token:

frun token        # prints a URL with a token in it — open that

Tokens are single-use and expire in about a minute. The page is handed its next one over the socket it already has, so a conversation keeps going without asking again. If you open the bare URL on a server with keys, the connection closes and the page says the token wasn't accepted.

Naming models

A model is a catalog id (frun models list), a file path, hf:owner/repo for anything on Hugging Face, or a URL for an OpenAI-compatible endpoint. Settings the config knows are applied; anything else is passed through to that runtime.

Secrets never go in the agent file — it names the variable holding a key (api_key_env="GROQ_API_KEY"), so agent.py is safe to commit.

On your own site

<script src="https://your-server/fusion-runtime.js"></script>
<button id="talk"></button>
<script>FusionRuntime.attach({ button: "#talk" });</script>

The runtime serves the browser client it uses itself, so the page you demo with is the one your site embeds. With no url it connects back to wherever the script came from.

Browsers only allow a microphone on https://, so a deployment needs TLS and wss://. A page never holds an API key: your backend mints it a short-lived token.

Authentication

frun key new
FUSION_ACCEPTED_KEYS=web:frun_kR7m...

Without keys the server answers on localhost only, and frun up --host 0.0.0.0 refuses to start. frun talk, a backend or curl send the key in an Authorization header; a browser page gets a short-lived, single-use token from POST /v1/sessions instead, because a page can hold neither a secret nor a header.

Concurrency caps, message and audio limits, idle timeouts, origin allowlists and proxy trust all have working defaults — see the docs.

Performance

Measured, not estimated. The production profile as it ships — RTX 3090, Qwen 7B q4 + Whisper small + Kokoro on the one card — through the browser client, 21 September 2026:

Two numbers, because there are two honest answers. A stopwatch started at your last syllable reads 991 ms. About 500 ms of that is the runtime waiting through silence to decide you've finished — which elapses while you're still finishing, so people don't experience it as waiting. What a caller feels is closer to the 490 ms of processing. Quote whichever you like, but say which one: a voice stack claiming a number under 500 ms is almost always measuring from "we decided the caller stopped", not "the caller stopped".

The stages don't sum, and that's not sleight of hand. Transcription of what you already said runs during the silence wait. And "first token → first audio" is mostly the language model writing a sentence — text-to-speech can't start on half a clause — so it is not a measure of how fast Kokoro is. Kokoro's own speed is the real-time factor: 0.09, or about 126 ms of compute for 1.4 seconds of speech.

Every figure is the runtime's own per-turn telemetry (frun talk --verbose, or the browser console), so you can reproduce them rather than trusting ours. Barge-in fired on every attempt.

Several callers at once

Measured on the same 3090, real WebSocket sessions, three turns each:

Four simultaneous callers land in the same range as one, within run-to-run variance. Past that it saturates: throughput plateaus around 0.7 turns/sec, so an extra caller past the knee buys queue time rather than capacity. Eight is not a conversation.

The bottleneck is one specific thing. At twelve callers the language model's first token takes 4790 ms of a 5312 ms response, while speech-to-text stays at 76 ms and text-to-speech at 469 ms. A single in-process llama.cpp context decodes one reply at a time; the speech stages do not care how many callers there are.

So to go past four, move the language model out and leave speech where it is:

llm = LLM("http://localhost:8080/v1", model_name="qwen2.5-7b-instruct")

vLLM and llama-server -np N both speak the API the openai_http runtime uses. Whether that moves the knee, and how far, is not yet measured.

The frun CLI

fusion-runtime works as an alias for frun.

Using it as a library

frun up agent.py covers running an agent. The pipeline can also run inside your own process — for a queue worker, a test, or a batch job over recorded calls — with no server involved. See examples/sdk_example.py, which is runnable, and the docs.

Documentation and contact

Everything else — configuration, turn detection, languages, the server API, telemetry, limits, GPU setup and deployment — is at fusion-runtime.dev/docs.

Development

uv sync --extra dev --extra talk     # or: pip install -e ".[dev,talk]"
pytest

CI runs the suite on Python 3.11, 3.12 and 3.13. CONTRIBUTING.md has the layout, the design rules a review will hold you to, and how to add a runtime.

License

Apache-2.0. Embed it in a commercial product, rebrand it, ship it closed — keep the copyright notice and the NOTICE file in what you distribute, and don't use the project's name to imply it endorses you.

The models it downloads by default are permissive too (Whisper MIT, Silero VAD MIT, Qwen2.5 Apache-2.0, Kokoro Apache-2.0), so the whole default path is clear for commercial use. A model you point it at yourself carries its own licence — check that one before you ship it.