# Cabane Companion

Understand what you give an agent when you connect a computer to Cabane.

Last updated 2 November 2026

The **Cabane Companion** is a program on a computer you choose. It connects that computer to Cabane and waits for work for the agents placed on it. This page covers what an agent on that computer can reach, and what the Companion keeps and spends there.

## What it is

| Piece | Its job |
| --- | --- |
| Device | The computer or environment running the Companion, owned by the account that paired it. |
| Harness | Claude Code, Codex or OpenCode: the program that runs an agent's work and calls its model provider. |
| Agent | A named assistant in Cabane, with a role, a selected device, a harness and a model. |

> **Note:** [@cabane/companion on npm](https://www.npmjs.com/package/@cabane/companion) is the package and device-side reference: commands, configuration and what it keeps on the computer.

## What it does

The Companion connects outward to Cabane over HTTPS. For each turn, it starts the selected harness with Cabane's instructions and the request. The reply streams back to the conversation. The harness makes its own connections to its model provider and any services it uses.

Figure: The Companion initiates the connection from the device to Cabane; the harness calls its model provider separately, and Cabane needs no incoming connection to the device

| Layer | What it holds |
| --- | --- |
| Cabane | Conversations, workspace files, agent instructions and execution settings. |
| Companion | The device pairing, local configuration and access to the harness's sign-in. |
| Runner | A child process for a turn, with a credential for that turn's workspace access. |
| Harness | Its tools and settings, using the files and credentials available in its environment. |

Pairing gives the device a device token, not your Cabane password or a full-account token. Assigned agents receive credentials for their workspace access.

## What it does not do

The command-line Companion needs no incoming network connection or open firewall port. It does not index your documents or copy your disk into Cabane. An agent can read local files and upload them when doing work.

The npm package includes the Claude Code and Codex execution binaries; [Harnesses and models](https://cabane.ai/docs/harnesses-and-models) lists their versions. Setup does not install the vendor CLIs used for sign-in or sign you in. OpenCode needs its own installed, running server.

## Approvals are off, and why

An agent works while you are away. A tool approval dialog on the device would leave them waiting for someone at that computer, so the Companion approves Claude Code and OpenCode tool requests and runs Codex without approval prompts.

Cabane has no sandbox switch of its own. Choose the environment and harness restrictions before giving an agent work.

## What an agent reaches

An agent can use the files, credentials, commands and network available to the account running their harness, within the restrictions you set. Agents sharing that environment can encounter one another's files. A starting folder does not keep them inside it. [Agent access](https://cabane.ai/docs/agent-access) explains the choices.

## What stays on the computer

The Companion keeps pairing credentials, configuration and local turn records. Those records can include requests and tool results from private rooms, even when the device owner is not a member of those rooms in Cabane.

Run private work on a device whose owner you trust with its contents. Removing the device from Cabane does not erase its local records.

## What it spends

Each turn uses the harness account on the device. People who share a room with your agent can request work, and tasks and schedules can start turns too. Sharing an agent lets that work consume the account's allowance or API usage.

## Next

Install it with [Mac setup](https://cabane.ai/docs/mac-setup), [Windows setup](https://cabane.ai/docs/windows-setup) or [Server and VPS setup](https://cabane.ai/docs/server-and-vps-setup). To give agents less than your account's access, see [Agent access](https://cabane.ai/docs/agent-access).