# Muse Code

Run Meta's terminal coding agent in Paseo with your existing Muse installation and credentials.

## Getting started

Install Muse Code on the machine running the Paseo daemon:

```bash
curl -fsSL https://dev.meta.ai/install.sh | sh
muse login
```

Use Muse 1.3.0 or newer. This integration was verified with 1.4.1. For API access, set
`META_API_KEY` in the daemon's environment instead of logging in. See
[provider environment overrides](/docs/custom-providers).

In Paseo, select **Muse Code**, then choose a model, approval mode, and reasoning effort.
Paseo runs `muse serve` and communicates over the Muse Session Protocol (MSP). Your Muse
configuration determines the available models.

## Provider options

Paseo launches Muse with the sandbox disabled (`--disable-sandbox`) and workspace trust
on (`--trust-workspace`). Muse can install packages using your npm cache and load project
rules, skills, and configuration.

Set defaults for every Muse agent in `config.json` on the daemon machine under
`agents.providers.muse.options`:

```json
{
  "agents": {
    "providers": {
      "muse": {
        "options": {
          "sandbox": { "enabled": true, "network": "proxy-only" },
          "trustWorkspace": false
        }
      }
    }
  }
}
```

You can also supply per-agent `providerOptions` when creating an agent. The daemon merges
provider defaults with the per-agent options and delivers the effective `providerOptions`
to Muse. See [provider options](/docs/sdk/provider-options) for configuration and creation.

Per-agent options are saved with the agent and reapplied when its Muse host opens on refresh
or resume. Invalid values and unknown keys fail session creation with a `providerOptions`
error.

| Option            | Default        | Effect                                                                                                                          |
| ----------------- | -------------- | ------------------------------------------------------------------------------------------------------------------------------- |
| `sandbox.enabled` | `false`        | Enables Muse's filesystem and network sandbox. Its read-only home directory prevents npm from writing the default cache.        |
| `sandbox.network` | `"proxy-only"` | Sandbox network policy: `"proxy-only"`, `"restricted"`, or `"enabled"`. With sandbox disabled, Muse allows full network access. |
| `trustWorkspace`  | `true`         | Lets Muse load project rules, skills, and configuration.                                                                        |

Approval modes control tool decisions separately: **Default**, **Ask**, **Strict**, and
**Full access**. Full access automatically allows escalated approval stages; it preserves
the effective provider options. Skills appear in the slash-command menu, alongside `/compact`.

## Limitations in Muse 1.4.1

- **Paseo MCP tools are unavailable.** Muse drops the returned tool namespace before matching
  an MCP call. This upstream issue must be fixed in Muse before its agents can use Paseo tools.
- **Compaction needs a model context limit.** If your endpoint does not report one, configure
  `context_compaction.provider_context_limit_tokens` in Muse for the selected model.
- **Usage windows appear only when Muse reports them.** An empty `usage/read` response
  leaves subscription usage unavailable.
- **Native subagents may need Muse's experimental SDK flag.** Set
  `MUSE_EXPERIMENTAL_SDK_ENABLED=on` in the provider environment if native subagent tools
  are missing. Linked child sessions appear only when Muse supplies `childSessionId`.

See [custom providers](/docs/custom-providers) to override the binary or environment under
`agents.providers.muse`.