The Agents tab

The agents a workspace can be launched on. Each carries its own credential, stored in the controller database and written into a workspace as it is provisioned.

Settings → Agents. A list beside an editor, the same shape as the seed files tab, because it is the same job: several configured things, one open at a time.

The fields

Several agents can share a Type. Two Claude subscriptions, each with its own credential and its own name, is the case this exists for. Type is which software drives the agent, not which agent it is.

Leave Model empty unless you mean it. Each agent then picks for itself and stays current as models change. A name pinned here is a name that goes stale with nothing to notice.

A disabled agent stays in this tab — that is how it gets re-enabled — and is absent from the launch form.

The credential

Write-only, and enforced by the shape rather than by a line someone could delete: the type the list endpoint returns has no credential field at all, and the query names its columns rather than selecting everything, so a secret column added later cannot quietly widen what the browser sees.

Blank on an edit means "keep the stored one". It is the one field where blank cannot mean empty. Otherwise renaming an agent would silently blank its token, and you would find out at the next provision.

The formats are not the same, because the agents are not:

Claude Code takes the string from claude setup-token. That needs a browser and a Claude subscription. It is not an API key.

opencode takes a JSON envelope naming the integration the value belongs to:

{ "integration": "openai", "value": { } }

An envelope rather than the bare value because the value does not say which integration it is for, and opencode stores the integration as a separate column. opencode has no API that accepts a credential — its OAuth routes need a person with a browser — so the runner writes a row into opencode's own database, which it reads on the next start.

Credentials are capped at 16 KB. Generously, on purpose: a Claude token is a few hundred bytes and opencode's about two thousand, and the cap is there to stop a paste accident becoming a database row rather than to express a belief about what a token looks like.

When a change takes effect

On the next workspace. A credential is written into a container at provision time, so:

• Rotating one leaves every running workspace alone. No restart, and nothing to redeploy.
• Deleting an agent leaves workspaces already running on it untouched. Their runner is installed and their credential is already in the container; nothing about them reads the row again.
• An agent that is up and unauthenticated cannot be rescued by editing its credential here. Fix the credential, then launch a new workspace.

This is the same split the rest of the controller uses: .env holds what the controller is and wants a restart; the database holds what it does and does not. See configuration versus policy.

Permissions

There is no permission-mode setting here, and there used to be. It was a second gate on a question each agent's own configuration already answers — Claude from .claude/settings.json, opencode from its permission lists — and two places deciding one thing, in two vocabularies that do not translate, is how they end up disagreeing.

Configure it per agent with a seed file instead. Whatever the agent does not resolve for itself still reaches the approval UI.