Self-host Agenta

A self-hosted deployment runs the studio, the API, the agent runner, and their datastores on your own infrastructure. It has every feature of Agenta Cloud except the governance features listed under Editions.

Start here

• Quick start runs the full stack locally with Docker Compose.
• Deploy on a remote server puts that stack on a server you own.
• Deploy to Kubernetes runs it from the Helm chart.
• Deploy on Railway is a community-maintained deployment. There is no official support for it, and fixes depend on contributor availability.

Where agents run

Agents execute in the runner service, in one of two sandboxes. Pick one before you open the deployment to other people:

• Run agents in a cloud sandbox (Daytona) gives every run its own isolated cloud sandbox. Use it for multi-user deployments.
• Run agents locally executes agents inside the runner container. Local runs are the default, and they are not isolated from each other, so use them on a deployment you run for yourself or for a trusted team. They are also the only runs that can use your own Claude, Pi, or ChatGPT subscription instead of a model API key.

Sandbox isolation and security covers the difference in full. Customize the agent runtime adds tools, folders, or compute to either one.

Editions

• OSS is the open-source edition, under the MIT license. It covers the studio, the API, the runner, and both sandbox providers. Every page in this section applies to it.
• EE is the enterprise edition. It adds the governance features: SSO, roles, and access controls. It is licensed separately.

Operate a deployment

• Configuration lists every platform environment variable.
• Runner and sandbox configuration lists every runner variable.
• Architecture and Networking describe the containers and the traffic between them.
• Upgrading covers version upgrades and migrations.