pve-agentsThe Agents tab
Configure the coding agents available for new workspaces.
Open Settings → Agents to create the agents operators can select at launch. Each configured agent has its own credential, stored in the controller database and written only into workspaces provisioned with it.
A workspace cannot launch until at least one enabled agent exists. The controller can otherwise be healthy, so add one before testing provisioning.
Fields
| Field | Meaning |
|---|---|
| Name | Operator-facing name shown on the launch form. |
| Type | Harness this release can run. See agents and harnesses. |
| Credential | Write-only credential installed during provisioning. |
| Model | Optional provider/model for opencode; ignored by Claude Code. |
| Enabled | Offers the agent for new workspaces. |
Several configured agents can share a type, such as two Claude subscriptions. Leave Model blank unless you need a specific version so the harness can use its current default.
Credentials
Credentials are never returned to the browser. When editing an agent, leave the credential blank to keep its existing value.
- Claude Code: the token from
claude setup-token; it is not an API key. - opencode: a JSON envelope that identifies its integration, for example:
{ "integration": "openai", "value": {} }Malformed credentials are rejected on save. Credentials are limited to 16 KB.
When changes apply
Changes affect future workspaces only. A running workspace retains the runner and credential written during provisioning. To recover an already-running agent with bad credentials, update the configured agent and launch a new workspace.
Configure agent-specific permissions with seed files. Requests the agent does not resolve itself appear in the browser approval flow.
Last updated on