Skip to content

Claude Code. Any Model. The most powerful AI coding agent now speaks every language.

8.2 KB
Raw Blame

Muse Code

Run Claude Code on Meta's Muse models with your Muse Code subscription.

Added in the first claudish release after 10.6.1. Last updated 2026-10-16.

muse@ signs requests with your Muse Code login and bills them to the subscription. meta@ is the separate, metered Meta Model API, signed by META_API_KEY.


First run

claudish login muse
claudish --model [email protected] "explain what src/index.ts does"

claudish login muse prints a code and opens Meta's sign-in page:

Sign in to Meta for Muse Code (a Muse Code subscription is required).

  Code: KQWD-PLRT
  URL:  https://auth.meta.com/oauth/device/?code=KQWD-PLRT

Approve the request in your browser; waiting…

Signed in to Muse Code (Muse Code Power Usage). Try it with: claudish --model [email protected] "hello"

The tier in parentheses is the one Meta reports for your account.

Prerequisites: a Muse Code subscription

  • An active Muse Code subscription on your Meta account
  • a claudish release newer than 10.6.1 (claudish --version)
  • For meta@ only: a Meta Model API key from https://dev.meta.ai

You do not need the Muse Code CLI. If you already use it, claudish can borrow its login instead.


Two ways to log in

claudish login muse the Muse Code CLI's muse login
Where the login lives ~/.claudish/meta-muse-oauth.json the macOS Keychain item ai.meta.dev.credentials
Works on any OS macOS only, for claudish's purposes
Dialogs none after the browser approval one macOS Keychain access dialog, the first time
Works in team slots and MCP sessions yes after one interactive approval (below)

claudish tries them in this order: its own login, then its copy of the Muse CLI's login, then the Keychain. claudish never writes or deletes the Muse CLI's Keychain item.

Borrowing the Muse CLI's login

Run one interactive launch in a terminal:

muse login                                   # once, in the Muse Code CLI
claudish --model [email protected] "hello"

Before Claude Code starts, macOS asks whether to allow access to the ai.meta.dev.credentials Keychain item. Click Always Allow. claudish then confirms the subscription with Meta and saves a copy of the login to ~/.claudish/meta-muse-borrowed.json (mode 0600).

Why team slots and MCP sessions need that copy

A team slot, a create_session pane, the MCP server, --probe, claudish serve and any piped or CI run never open a Keychain dialog. Nobody watches their screen, so an unanswered dialog would block the request. These processes read only claudish's two files, and without either one they stop with:

muse@ in a team slot / MCP session needs a claudish login: run `claudish login muse`, or one
interactive `claudish --model muse@<model>` and allow Keychain access.

Either fix works, and from then on every slot signs on its own with no dialog.


Choosing a model

Today you name the model with its exact wire id:

claudish --model [email protected] "refactor the auth module"
claudish --model [email protected] "review this diff"
claudish --model [email protected] "write tests for utils.ts"

Run claudish with no --model to open the picker. Its Muse Code row lists the models your account serves, read live from Meta (/muse-code/models).

Bare names do not reach muse@ yet. The cloud models catalog has no route for the Muse Code subscription. Until it does, a bare name such as muse-spark-1.3 goes to meta@ if META_API_KEY is set, then OpenCode Zen, then OpenRouter. Check what a name resolves to:

claudish --probe muse-spark-1.3

A bare name leads with muse@ once the catalog publishes that route.

Contributor models

Each Muse chat model also has a -contributor twin (muse-spark-1.3-contributor). Meta's description for those ids: "Your content, including inter-session messages, may be used for product improvement." claudish never picks a contributor id for you, on any provider. It sends one only when you type it exactly:

claudish --model [email protected] "hello"

Reasoning effort

claudish sends the effort level Claude Code asks for. It encodes that level the way the metered Meta API's catalog entry for the same model does.

The Wonder Full model is the exception. The metered API has no entry for muse-wonder-full-1.0, so claudish cannot encode its effort, and the request carries the Responses adapter's default. Contributor ids take only the output limit from their entry, with no effort encoding.


Billing: muse@ and meta@

Metered meta@

export META_API_KEY='your-meta-model-api-key'
claudish --model [email protected] "summarise CHANGELOG.md"

Why two providers? META_API_KEY never signs muse@, and your Muse Code login never signs meta@, because each credential is billed differently.

meta@ used to be an OllamaCloud shortcut. It now means Meta. Use oc@, llama@ or lc@ for OllamaCloud.

Who bills what

Provider Billed claudish shows
muse@ your Muse Code subscription, flat rate SUB, $0 per request
meta@ per token, to your Meta Model API account the per-token price

Before it signs any muse@ request, claudish asks Meta's key service whether your subscription is active and needs no payment method. With no confirmation it does not sign. It asks again every 30 minutes, so a lapsed subscription stops within 30 minutes even in a long session.

If your Muse Code login expires mid-session, the request stops with a login message. claudish does not move the request to a metered provider.

Where your login goes

claudish sends your Muse Code login, the key it mints from it, and every muse@ request only to https://api.meta.ai. There is no setting to change that host: a .env file in a project folder could otherwise redirect your login to someone else's server.

Logging out

claudish logout muse

This deletes ~/.claudish/meta-muse-oauth.json and ~/.claudish/meta-muse-borrowed.json. It leaves the Muse Code CLI's own login alone, and it does not revoke the token at Meta. To end the Muse CLI's session too, use the Muse Code CLI's own logout.


Troubleshooting

"muse@ needs a Muse Code login"

Cause: no claudish login and no borrowed copy, and no Muse CLI login in the Keychain.

Fix: run claudish login muse.

"Your Muse Code login has expired"

Cause: Meta's key service rejected the login (HTTP 401 or 403). Logging out of the Muse Code CLI revokes the token claudish copied.

Fix: run claudish login muse, or muse login followed by one interactive claudish launch.

"Your Meta account has no active Muse Code subscription (or needs a payment method)"

Cause: the login works, but Meta reports no active subscription. Logging in again does not change this.

Fix: subscribe or add a payment method at the URL in the message (https://accountscenter.meta.com/muse_code/?ep=no_payg when Meta names no other).

"could not read the Muse Code login from the macOS Keychain"

Cause: the access dialog was denied, or got no answer within 120 seconds.

Fix: run the launch again and click Always Allow, or use claudish login muse, which needs no Keychain access. claudish never kills the waiting Keychain read. A late answer still counts, and only one dialog is open per machine at a time.

"Muse Code key service unavailable (HTTP 503)"

Cause: Meta's key service is down or rate-limited (HTTP 429 or 5xx).

Fix: retry. Claude Code's own retry re-sends to muse@; the request does not move to another provider.

"muse@ in a team slot / MCP session needs a claudish login"

Cause: you logged in with muse login only, and no interactive claudish launch has copied that login yet. The slot may not read the Keychain itself. See Why team slots and MCP sessions need that copy.

Fix: run claudish login muse once, or one interactive claudish --model [email protected].


Next steps