Skip to main content
OpenComputer keeps the agent development experience code-first while managing the infrastructure required to run it. You work with one public surface—the opencomputer CLI—and the same agent project can run locally or in a managed sandbox.

The layers

Agent project

The project is the portable definition of the agent. Its committed files describe:
  • stable UUID and readable display name in opencomputer.toml
  • behavior and approval rules in instructions.md
  • runtime and model configuration
  • tools and skills
  • required connections and channels
  • initial workspace files and evaluations
The project contains declarations and code, not provider credentials.

OpenComputer CLI

The CLI owns the developer workflow:
  • discover templates
  • initialize a complete project
  • authenticate the developer
  • authorize managed connections
  • run the project locally with OpenCode
  • package and deploy the project
  • start sessions from deployed agents
When you run opencomputer deploy, the CLI builds a canonical source artifact from the agent directory and sends it to OpenComputer. Infrastructure provider details are not part of the CLI contract.

Agent identity and deployments

The UUID in opencomputer.toml identifies the agent across its lifetime. Its two-word name is display metadata and can change without creating a different agent. Each distinct source artifact becomes an immutable deployment.
Aliases such as production provide a stable target while preserving the exact source behind every deployment. Deploying unchanged source is idempotent; deploying changed source creates a new version.

Managed sessions and sandboxes

A deployed run creates a managed session from the selected agent version. OpenComputer prepares an isolated sandbox, loads the agent source and workspace, starts OpenCode, and streams the result back to the CLI. The sandbox is an implementation detail of the agent experience. You do not need to create or manage a separate sandbox before deploying an agent. Each session has:
  • the exact deployed source version
  • an isolated filesystem and process environment
  • a workspace that the agent can use during the task
  • access to only the connections and channels declared for that agent
  • lifecycle management handled by OpenComputer

Connections

Connections let agent tools act on services such as Gmail without committing OAuth tokens to Git. For Gmail:
1

Declare

The template places a Google connection declaration in connections/google.json.
2

Authorize

opencomputer connection add gmail --alias personal opens the account authorization flow. Add another alias to connect another Gmail account.
3

Develop locally

The CLI gives the local agent a short-lived, scoped route to the managed connection.
4

Run when deployed

The managed session receives the same connection capability inside its sandbox. Provider credentials remain outside the project and agent output.
An agent’s instructions should still define explicit approval boundaries. Connection scope controls which service operations are possible; instructions control when the agent should use them. The Gmail connection is centralized for the OpenComputer user. Local development does not create a second local OAuth installation:
The repository contains only the connection declaration. OAuth credentials stay outside Git and are never given directly to the model.

Connections and aliases

Learn how connection declarations, user identity, and multiple accounts work.

Local and deployed execution

This symmetry is intentional: develop the actual agent locally, then deploy the same directory instead of rebuilding the workflow in a dashboard. Managed sessions can be listed, inspected, attached to, resumed with another turn, and ended from the CLI. A session is the durable unit of interaction; the sandbox runtime may suspend between turns and resume when needed.

Deployment sequence

Security boundaries

  • Agent source and identity are reviewable and commit-friendly.
  • OAuth credentials are managed separately from source code.
  • Local connection access uses a scoped session capability.
  • Deployed agents run inside isolated sandboxes.
  • The public CLI and API do not expose infrastructure-provider credentials or storage implementation details.
  • Consequential actions should remain behind explicit instructions and user approval.

Deploy a Gmail summarizer

Follow the complete CLI workflow.