The source-available runtime for business brains, built and used through your agent.
Save the way your team reviews a campaign brief as a reason function. An agent supplies a brief, runs the review and reads the recorded result. The definition stays in the brain, ready for the next brief or another authorized colleague.
- Add your Auto Cloud workspace's MCP URL and authentication details to your agent. Need access? Contact Auto. See connection instructions or request a Studio invite.
- Paste this into your connected agent:
Create a brain named Quickstart with id quickstart, or reuse it if it exists.
List the available models and ask me to choose a concrete model or alias.
In that brain, create a reason function called check-brief. It accepts a brief
as text and checks for an audience, a budget and a measurable goal.
If the function already exists, show it to me before making changes.
Run the saved function on: "Promote our reporting tool to finance teams
with a USD 5,000 budget."
Show the result and read back the recorded run, including its execution id.
The review should flag the missing measurable goal. You now have a saved function you can reuse and a recorded run to inspect. Your connection needs permission to create brains, create and run functions, and use a model. If model discovery is unavailable, ask your workspace administrator for a concrete model reference.
For a second run and a comparison of the results, continue with Build your first brain.
Run locally instead
Install pnpm, then:
git clone https://github.com/BeOnAuto/auto-brain.git
cd auto-brain
pnpm install
cp .env.example .envIn .env, set a model provider key such as ANTHROPIC_API_KEY, OPENAI_API_KEY or GOOGLE_GENERATIVE_AI_API_KEY. Keep it out of your agent conversation. pnpm uses the Node and pnpm versions pinned by this repository.
pnpm dev:leanKeep that terminal open. The runtime listens at http://localhost:8080; in another terminal, curl http://localhost:8080/health should return {"status":"ok"}.
Add http://localhost:8080/mcp as a streamable HTTP connection in an agent running on your computer, without an authentication header, then use the prompt above. A hosted agent cannot reach your laptop's localhost. This local mode has no authentication: do not expose it through a tunnel or public proxy. The lean server supports this reason-function example without starting workflows.
See local setup and client configurations for details. For a deployment your team operates, see Self-hosting.
Auto documentation covers the platform and this runtime together. Start with Brains and methods; consult Functions and availability for current capabilities, or the MCP reference when integrating an agent.
The source-available runtime is in early development and is not ready for production use; availability describes the current scope.
Coming soon: tool access inside reason functions through a shared catalog and MCP gateway, with direct tool lists and bounded tool-call loops. Today, the external agent calls Auto over MCP and passes evidence into the function; the function does not inherit the agent's tools. See Tool access.
For bugs and questions, open an issue. Report vulnerabilities through SECURITY.md.
Runtime code and its documentation live here. See CONTRIBUTING.md for development setup and checks, and Documentation contributions to preview or improve a guide. Documentation pull requests can be reviewed without the private website repository or service credentials.
auto-brain is source-available under the Elastic License 2.0. LICENSING.md explains the terms, including restrictions on offering it as a hosted or managed service. Contributions follow the CLA and Code of Conduct.