The open-source runtime that runs a whole AI company.
Not an agent framework you assemble into something useful. A running company — roles, a backlog, approvals, and agents that hand work to each other — that installs on your machine and starts without an API key.
- Open-core
- No API key required
- Your model, local or hosted
- No trial clock
Free means free. There's no clock.
The Community Runtime is not the 7-day hosted trial. It's the open core, it runs on your hardware, and it doesn't expire, phone home for a licence, or degrade into a nag screen. If you never talk to us again, it keeps working.
What we kept out of it: nothing that would make it useless on purpose. A crippled open core is a bad advertisement for the paid one.
One disclosure, so you don't have to find it yourself: first-party usage telemetry — event counts, never content — is on by default, and one environment variable turns it off.
You don't need an API key to see it work.
Most agent tooling assumes you'll arrive with a paid key and a billing account. BrainOutput ships with a free model default, so the first run costs you nothing — no API key, no account, no card. When you want more capability, point it at a key — or at a model on your own machine, where your prompts never leave your network.
Designed to put a first real result in front of you in about five minutes.
- Run AI agents free, without an API key (coming soon)
Your provider. Your key. Your call.
Claude · OpenAI · Gemini · OpenRouter · DeepSeek · OpenCode — and local models through Ollama and friends. Provider independence isn't a feature we bolted on; it's the reason the runtime exists. You are never locked to one vendor's pricing, one vendor's rate limits, or one vendor's opinion about what your agents may do.
Self-hosted, all the way down.
The whole company runs on infrastructure you control. With local models, your prompts, your work products and your customers' data stay on your side of the network boundary. For teams under EU data-residency obligations, that isn't a preference — it's the requirement.
- Which models fit my hardware? →
- Run it with Ollama (coming soon)
A company, not a crew.
Agent frameworks give you primitives and a loop. You still design the org, the hand-offs, the review step, and the place a human says yes or no. BrainOutput starts one level up:
- An org chart, not a graph. Roles with responsibilities, and a chain of command.
- A backlog, not a prompt. Work is issues that get picked up, worked, and closed.
- Review between agents. Work gets checked by someone other than its author before it reaches you.
- Approval gates you hold. Consequential actions wait for a human.
- Identity per member. Designed so each role has its own email address.
Judge it on the diff.
An agent took a ticket against a real Python repository it had never seen, followed the conventions already there, and opened a pull request. We cloned the pushed branch from GitHub into a clean directory and ran the project's own suite: 90 passing, up from 35. It's open for review, not merged, and the repo has no CI — so the verification is ours, and you can repeat it.
And it isn't only code. Lumière Café: someone described an independent coffee roastery in one paragraph. The agents returned a finished single-page site: semantic HTML, responsive, and zero external dependencies — no CDN, no fonts fetched, no scripts phoning home. It opens from a file, offline, with nothing installed.
Provenance: real runs on the Paperclip execution core, 2026-07-19 and 2026-07-17. The site's HTML is byte-for-byte as delivered.