openrtc

Python 3.11 to 3.13

Docs · Quick start · Benchmark · Contributing

from openrtc import AgentPool
from my_agents import BookingAgent, SupportAgent, SalesAgent

pool = AgentPool()                 # one livekit-agents worker
pool.add("booking", BookingAgent)  # plain livekit.agents.Agent subclasses
pool.add("support", SupportAgent)
pool.add("sales", SalesAgent)
pool.run()                         # each call is routed to one of them

OpenRTC is the ops layer for self-hosted LiveKit Agents: routing, hot reload, a live openrtc top, tenant caps and drain deploys, around the livekit-agents worker you already run.

Why

The problem. A livekit-agents worker has one entrypoint. With more than one agent you either write the dispatch yourself or run a worker per agent, and the tooling to operate it (which call is slow, which client is using the box, how to ship without cutting calls) is yours to build too.
What you get. One pool that routes each call to the right agent, swaps live calls to your edited code, shows every call in openrtc top, caps each tenant, and drains on deploy.
What doesn't change. Your agents stay standard Agent subclasses. No base class, and @function_tool, RunContext, on_enter and the *_node hooks are untouched. You delete the per-worker entrypoint and AgentSession wiring.

Quick start

pip install "openrtc[livekit,cli]"

Python 3.11 to 3.13, livekit-agents >=1.5,<1.9. Set LIVEKIT_URL, LIVEKIT_API_KEY and LIVEKIT_API_SECRET as for any LiveKit worker. Put one agent per file in a directory:

# agents/booking.py
from livekit.agents import Agent
from openrtc import agent_config


@agent_config(name="booking", greeting="Welcome to reservations.")
class BookingAgent(Agent):
    def __init__(self) -> None:
        super().__init__(instructions="You help callers book a table.")

openrtc dev ./agents \
  --default-stt openai/gpt-4o-mini-transcribe \
  --default-llm openai/gpt-4.1-mini \
  --default-tts openai/gpt-4o-mini-tts

openrtc top          # in a second terminal: every live call

A room named booking-call-1 now reaches BookingAgent. Edit the file during a call and the next turn runs your change. Prefer Python to a directory? pool.add(...) as above, then python main.py dev. Full quick start.

Using a coding agent (Claude Code, Cursor, Codex)? npx skills add mahimailabs/openrtc-runtime gives it two skills: moving a livekit-agents project onto OpenRTC, and running it in production.

What you get, and where it stops

	What it does	Limits
Routing	Each call goes to one agent by job metadata, room metadata, or room-name prefix.	An unknown agent name raises; routing never falls back silently.
Hot reload	`openrtc dev` moves live calls to your edited agent on their next turn. A bad save keeps the old code.	Coroutine mode only.
`openrtc top`	A live table of every call: agent, tenant, duration, CPU and memory; `slow` when a call blocks the others.	Same host. `slow` needs coroutine mode.
Tenants	Per-tenant providers and keys, per-tenant and per-agent caps, a circuit breaker for a failing tenant.	Not a sandbox. The breaker needs coroutine mode.
Deploys	`deployment_version`, `pool.begin_drain()` and audit events around livekit's drain.	Calls finish on the old version; none is moved.
Two isolation modes	`coroutine` (default): every call in one process. `process`: livekit's process per call.	Coroutine mode leans on livekit internals, hence the tight version pin.

Measured, not claimed

Against livekit-agents 1.8.3 on the same two cores, 8 calls: OpenRTC's default mode used about 22 MB per call against about 65 MB, at about the same CPU (uvloop). CPU, not memory, is what limits calls per machine, so OpenRTC does not make one box serve more calls. Benchmark, with the harness to rerun it in benchmarks/headtohead.

When not to use it: you have one agent and don't need the ops tooling (plain livekit-agents is simpler), or you need hard per-call isolation and hot reload together (isolation="process" isolates each call but cannot hot reload).

Docs

Why OpenRTC · How it works · CLI · Benchmark · Changelog · for AI coding tools: llms.txt

Contributing

OpenRTC is early, and a small change moves it. Three commands to a green run (make ci itself takes about a minute):

git clone https://github.com/mahimailabs/openrtc-runtime && cd openrtc-runtime
uv sync --group dev
make ci          # ruff, format, mypy --strict, pytest with the 99% coverage gate

Where to start: issues labelled good first issue are scoped, with the files to read and what done looks like. help wanted are bigger. Found a bug or have an idea? Open an issue.

How we work: read CONTRIBUTING.md. Performance claims need a measurement. AI-assisted PRs are welcome: the repo ships a CLAUDE.md and agent skills (skills/ for adopters), and a PR is judged on the same make ci and review either way.

License

MIT. Built by Mahimai Raja at Mahimai AI, in public, on LiveKit Agents.

Metadata

Release files for openrtc 0.20.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for openrtc 0.20.1
File	Size	Uploaded
openrtc-0.20.1.tar.gz	1.8 MB	Sep 27, 2026	Details

Built distribution (wheel)

Table of built distributions (wheels) for openrtc 0.20.1
File	Interpreter	ABI	Platform	Reset
openrtc-0.20.1-py3-none-any.whl 139.7 kB Sep 27, 2026	Python 3	none	any	Details

Total release size: 1.9 MB

Release files / openrtc-0.20.1.tar.gz

Download URL	openrtc-0.20.1.tar.gz
Size	1.8 MB
Tags	Source
SHA-256 checksum How to use checksums	`d1fabdb0fd7e3d692ff26d742d7a430bc740745852a12e98d157db58f7739db1`
BLAKE2b-256 checksum How to use checksums	`52d65c473a97d1091c62fe1d213c5fe31235ad69f3c77829f7e446eb12fd5266`
Upload date	Sep 27, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.13.14`

Release files / openrtc-0.20.1-py3-none-any.whl

Download URL	openrtc-0.20.1-py3-none-any.whl
Size	139.7 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`a5559999238fad58eb04c7fabb61283a82a4af1ef1c6d6f455a5634c745b5039`
BLAKE2b-256 checksum How to use checksums	`d2bb44b6da08e9210544136f5ff5740df716481f43c5be0e6d7095916fb73fdf`
Upload date	Sep 27, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.13.14`