Architecture
RunMat Server is a single service that handles sign-in, organizations and projects, project files and history, live collaboration, the package registry, the agent's models and remote execution. Its state lives in PostgreSQL and an object store, so server instances can be replaced or added without moving data.
This page lists the components a deployment needs and what each one is for. Networking covers what needs to reach what. Enterprise customers receive the full deployment guide with their release kit.
Components
Loading diagram...
| Component | Required | Role |
|---|---|---|
| RunMat Server | Yes | The RunMat API. Instances keep no state of their own, so you can run as many as you need behind a load balancer. |
| PostgreSQL | Yes | Accounts, organizations, projects, file history and usage. The server applies its own schema updates at start-up. |
| Object store | Yes | File content, datasets, run artifacts and packages. S3-compatible storage or Google Cloud Storage. |
| Identity provider | Yes | Signs users in. See Sign-in, SSO and SCIM. |
| Redis | For live collaboration, or more than one server instance | Short-lived coordination between instances. It holds nothing that needs a backup. |
| Search cluster | Optional | Search across project files and the documentation. Elasticsearch-compatible. |
| Model provider or gateway | Optional | The agent's language models. |
| Compute nodes | Optional | Machines that run remote and cluster jobs. |
Desktop and the CLI run code on the user's computer by default. A server adds shared projects, history, live collaboration, packages, the agent's hosted models and remote execution.
What needs to reach what
- Clients and compute nodes reach the server over HTTPS. The load balancer must pass WebSocket and Server-Sent Events connections through, because live collaboration and change notifications use them.
- Clients also reach the object store directly: large uploads and downloads go there over short-lived signed URLs that the server issues.
- The server reaches PostgreSQL, the object store, Redis and the identity provider, plus the optional services you enable.
See Networking.
Data and security boundaries
- PostgreSQL and the object store together hold everything durable. Back them up as one set.
- The server is the source of truth for accounts, membership and committed file versions. Clients cache files, but a cached copy never overrides the server's version.
- Remote runs are encrypted on the user's computer for the node that runs them. The server schedules and relays them without being able to read the program or its data. Local runs never pass through the server.