Cosmo Platform

Build your own Cosmo.

The same architecture that makes Cosmo a partner — long-term memory, realtime presence, and continual learning — behind one API.

Define

Model, system prompt, tools.

Deploy

Full-duplex, sub-second, one API.

Grow

Learn from every trace, with your gates.

Generate API Key

1from cosmo import Cosmo
2 
3cosmo = Cosmo(api_key="sk-...")
4 
5# Define a realtime agent. Cosmo runs the transport and the inference.
6agent = cosmo.agents.create(
7    model="cosmo-realtime",
8    system=SUPPORT_AGENT_PROMPT,
9    tools=[get_account, issue_refund],
10)
11 
12# Start a live session.
13async with agent.start() as session:
14    async for event in session:
15        ...  # agent converses; observe transcripts
16 
17# Every session becomes a replayable trace. Cosmo grows the agent
18# and ships an update only past your regression gate.
19agent.improve(care_about="suites/refunds.jsonl")

Why Cosmo

The realtime stack solved the pipes. Not the smarts.

Frameworks like LiveKit and Pipecat standardized transport, but an agent’s latency and cost live in the models, not the network. And the moment you make one improve from production, you hit the wall no framework addresses. An update helps on average and quietly breaks the one flow you can’t afford to lose: the refund it shouldn’t approve, the escalation it stops making.

Cosmo is built for that wall. One runtime for the transport and the inference, where improving the model is routine instead of a risk.

Without Cosmo

Static agent — never improves after launch

Fine-tune by hand: better on average, broken somewhere you didn’t test

Quality and latency invisible until a user complains

Frontier-model cost on every turn

With Cosmo

Learns from every live interaction

Every update clears a hard regression gate

Continuous, replayable measurement of both

Specialized serving at a fraction of the cost

One runtime beneath your agent.

Cosmo’s runtime hosts the transport, from full-duplex audio and video to SIP, and the model inference. Every session is captured as a replayable trace and used to improve the model, under the gate you define. You build the agent. We run everything beneath it.

Your agent · what you build

system prompttoolsbusiness logic

↕

Cosmo · realtime runtime

full-duplex transportmodel inferencecontinual learningthe gate

↕

Model providers & accelerators

open weightsfrontier APIsyour GPUs

Team

Built by people who’ve run this problem at scale.

We’ve served low-latency models to billions of users and kept Messenger’s stateful sync layer running. We’ve also run continual-learning loops where a single regression was treated like an outage.

Google · Meta AIResearch at Google (Gemini) and Meta AI across voice, RL, reasoning and continual learning.
Meta · IntegrityIntegrity classifiers at company scale: low-latency inference in a continual-learning loop where regressing was the cardinal sin.
Meta · MessengerMessenger sync infrastructure. Stateful, full-duplex realtime communication at billion-user scale.
VMware · MetaSystems and runtime engineering across VMware and Meta.
Microsoft · Rubrik · Stripe · Oracle · And More

Multimodal, long-horizon, self-improving agents

Start in a few lines, or talk to the team about your realtime workload.

Get started