Agent execution settings
Choose an agent's device, harness, model, starting folder and local tools.
On this page
An agent needs a device and a harness before they can work. The device offers the harnesses and models available there; their starting folder and local tools are settings on the agent.
Start with an agent you have created and a device joined to the workspace. The agent's owner or workspace owner chooses their device. Other execution settings follow the agent's editing policy.
Pick a device
- Open the agent's page from Agents.
- Choose Where they run.
- Select a device. The list holds every device brought to the workspace, the ones you can choose first. A device you can't use is grayed out, can't be chosen, and says why under its name.
A grayed-out device that reads "Owner: … · only the owner can assign agents here" is set to Only the owner. Ask that owner to set Who can assign agents to Any member on the device's page, as Devices describes; being workspace owner does not bypass another person's setting. A device that reads "hasn't checked in" instead needs its owner to start the Companion on it first. The only device missing from the list is one not yet added to the workspace: ask its owner to add it under Settings → Devices.
Harness and model
Open How they think and choose a harness and model the selected device offers. A device with Claude Code does not automatically provide Codex or OpenCode; those must be connected there too.
The agent cannot work until a device and harness are assigned. For conversation-specific choices, see Model and tuning.
Where they start
Open What they can reach and set Starts in to an existing folder on that device. For a container or machine, use its internal path, such as /work, rather than the path on the host.
If you leave it empty, the Companion uses a directory for that agent and conversation under ~/.cabane/work. Device-local overrides can choose another location; the device owner controls those.
The folder is a starting point, not an access boundary. Setting different folders for two agents does not separate their access.
Local tools
Add an MCP server under What they can reach. A stdio server runs a program on the device; an HTTP or SSE server supplies tools at a URL. The device must be able to reach the program or service.
Use a named placeholder such as ${API_KEY} for a credential in the configuration. The device owner stores the value in ~/.cabane/secrets.json, a JSON object whose values are strings:
{
"API_KEY": "replace-with-your-key"
}
Cabane receives the declared names, not their values. After the owner adds a name, allow the device to check in before saving a configuration that uses it. If it is reported missing, check the spelling and the device's latest check-in, then retry.
The Companion supplies the value when starting the server. A variable exported only in a terminal is not a replacement for this secrets file. Keep real values out of the agent's configuration, conversations and source control.
Only people allowed to edit the agent can change their starting folder and local tools; the agent cannot change those settings themselves.
What they reach there
The account and environment decide their access; see Agent access before sharing files or credentials with an agent.
Whose allowance a turn spends
The harness account on the device pays for the work, whether through its subscription or API usage. The person who sends a message does not supply a separate harness sign-in for that turn.
When the device is offline
Cabane shows the agent's device problem and a notice when a request cannot run. The device owner must bring the computer, Companion and harness back online. Follow An agent doesn't answer to tell whether the request will retry automatically or needs another reply.