# Muse Code — Meta's terminal and CI coding agent, built on the Muse Spark model.

> Source: The Agents Index — https://theagentsindex.com/muse-code (structured, researched, re-verified)
> Facts last verified: 2026-09-20

Muse Code is Meta's first-party coding agent for the terminal and CI, built on the Muse Spark model and installed as a native `muse` binary. It plans, edits files and runs shell commands inside an OS-enforced sandbox with approval prompts on by default.

| Fact | Value |
| --- | --- |
| Website | https://dev.meta.ai |
| Price | from $5/month |
| Free tier | Not applicable |

Three published plans, $5, $15 and $50 per month; no free tier. Pay-as-you-go on Model API is the alternative at $1.25/1M input and $4.25/1M output.

## What it does

**Writes code** · Yes — Authors and edits code from a plain-language instruction in a project directory, running as an interactive terminal session or a single headless prompt.

> Run it in a project and it plans, edits, and runs commands to do a task, with approvals and an OS sandbox on from the first run.
> — https://dev.meta.ai/docs/muse-code · Muse Code overview, opening paragraph

**Debugs and fixes** · Yes — The vendor positions debugging as a core job alongside building, and the agent can run test suites and summarise failures headlessly.

> A coding agent for your most complex coding workstreams. Build, debug and ship with Muse Code.
> — https://dev.meta.ai/lp/muse-code · Hero section, product subheading

**Reviews code** · Partly — Review happens inside its own runs — background reviewer and verification agents check the work — rather than as a review of a diff someone else wrote in a pull request.

> Multiple agents coordinate on every task. Workers in parallel, reviewers in the background, so you ship faster without sacrificing quality.
> — https://dev.meta.ai/lp/muse-code · Accelerate your build section, 'Multi-agent by default' card

**Whole-codebase work** · Partly — It splits a large job across parallel subagents with optional Git worktree isolation, which fits a repository-wide migration, but the vendor documents no repository indexing and caps a tree at eight concurrent agents by default.

> Distribute a large job across a team of agents that work in parallel, and steer them from one place. A lead session spawns child agents, hands each a bounded task, and keeps managing the group while they run.
> — https://dev.meta.ai/docs/muse-code/extending · Subagents and multi-agent section, opening paragraph

**Ships it** · Partly — It can run shell commands including git operations when you ask, but by design it stops short of committing or pushing on its own, and no deployment surface is documented.

> it won't commit, amend, or push your work unless you ask it to in the session, it keeps scratch files outside your checkout, and it reverts incidental edits it didn't need
> — https://dev.meta.ai/docs/muse-code/configuration · Project instructions with AGENTS.md, 'Conservative by default' note

**Done-for-you service** · No — This is self-serve software you install and run yourself with a one-line installer; Meta staff do not build or operate the agent for you.

> Install Muse Code with the one-line installer. It puts a native `muse` binary on your path.
> — https://dev.meta.ai/docs/muse-code · Install section

_Not established: Writes tests_

## How much it does unattended

**Runs autonomously** · Yes — A headless run takes one prompt to completion with approvals disabled, and session goals keep the runtime queueing follow-up work until the objective is met.

> `muse exec` takes one prompt, runs it to completion, and exits, so you can drive the agent from a script, a job, or a CI pipeline
> — https://dev.meta.ai/docs/muse-code/extending · Headless and CI section

**Multi-agent** · Yes — A lead session spawns child agents that can themselves spawn grandchildren, with eight concurrent agents per tree by default and a configurable ceiling of 64.

> One agent tree can execute eight agents at once by default, including the root agent. Set `agents.execution_capacity` from 1 through 64 in `settings.json` to change the limit.
> — https://dev.meta.ai/docs/muse-code/extending · Subagents and multi-agent section, properties list

**Agent permissions** · Yes — Four permission profiles from read-only to unrestricted, three approval modes, stage-by-stage shell review and per-destination network approval decide what runs before it asks.

> Approval and sandboxing are on by default. The agent asks before any risky action, and every shell command runs inside an OS-enforced sandbox.
> — https://dev.meta.ai/docs/muse-code/permissions · Permissions and safety, opening paragraph

## Which models it runs on

**Claude** · No — The plan table and the model-ID pricing enumerate only Muse Spark; no third-party model is offered inside Muse Code.

> Pricing is determined by model ID, so developers can pick the right tier based on their needs: model to muse-spark-1.2-contributor ... model to muse-spark-1.2: standard Model API pricing
> — https://developer.meta.com/ai/resources/blog/build-with-muse-code/ · Meet Muse Spark 1.2 section, pricing-tier enumeration

**GPT** · No — Same closed model enumeration: only the two Muse Spark 1.2 model IDs are selectable; OpenAI compatibility is SDK shape, not an OpenAI model.

> Point your existing OpenAI SDK compatible client at Meta Model API.
> — https://developer.meta.com/ai/products/muse-code/ · Cookbooks and quickstarts, Quick start card

**Open models** · No — Muse Code runs Muse Spark, a hosted proprietary model; the vendor's open-weight model, Muse Glimmer, is offered on Model API rather than in Muse Code.

> Muse Code is a fast and accessible coding agent powered by Muse Spark.
> — https://developer.meta.com/ai/products/muse-code/ · Accelerate your build section, intro line

**Model choice** · Partly — You pick among Muse Spark versions per run or mid-session, and the reasoning docs imply other providers exist, but no alternative provider is named or configured anywhere in the Muse Code docs.

> The default model is `muse-spark-1.2`. Override it per run with `--model`, or switch mid-session with the `/models` slash command
> — https://dev.meta.ai/docs/muse-code/configuration · Select a model section

## Where you use it

**In your editor** · No — The vendor describes exactly two run surfaces, interactive terminal and headless, with no editor or IDE extension mentioned anywhere in the Muse Code documentation set.

> The same binary runs two ways: - **Interactive**: `muse` opens a terminal UI for a back-and-forth session with slash commands, approvals, and live status.
> — https://dev.meta.ai/docs/muse-code · Interactive or headless section

**On the command line** · Yes — The terminal is the primary surface: a native binary with a TUI, slash commands, live approvals and a command palette on macOS, Linux and Windows.

> **Interactive**: `muse` opens a terminal UI for a back-and-forth session with slash commands, approvals, and live status.
> — https://dev.meta.ai/docs/muse-code · Interactive or headless section, first bullet

**In your pipeline** · Partly — It runs unattended in a pipeline with an API key and emits JSONL events, but the sandbox needs working bubblewrap on Linux and no pull-request integration is documented.

> The [sandbox] must be able to run on the CI runner, which on Linux needs a working bubblewrap and a non-musl build. Without it, every sandboxed shell command aborts as an environment failure.
> — https://dev.meta.ai/docs/muse-code/extending · Headless and CI section, 'CI sandbox requirements' note

**In a browser** · No — Muse Code itself has no hosted web surface; the dev.meta.ai dashboard issues keys and manages billing for the Model API rather than running agent sessions.

> Muse Code is Meta's coding agent for the terminal and CI, built for [Muse Spark](https://dev.meta.ai/docs/models#muse-spark).
> — https://dev.meta.ai/docs/muse-code · Muse Code overview, first sentence

## Whose machine it runs on

**Self-hosted** · No — The binary runs locally but every turn calls Meta's hosted model and refuses to start without a Meta credential, so inference cannot be run on your own infrastructure.

> Muse Code needs a Meta credential before it can call the model.
> — https://dev.meta.ai/docs/muse-code/auth · Authentication and billing, opening sentence

**Open source** · Partly — The client SDK and the MSP protocol declarations are MIT-licensed and public; the muse CLI itself ships as a prebuilt native binary with no published source.

> Released under the MIT License — see LICENSE.
> — https://github.com/meta-models/muse-code-sdk · README, License section

## What it costs to run

**How it meters** · Yes — Two metering models: pay-as-you-go per input, cached and output token, or a flat monthly subscription tied to the CLI's own key with prompt allowances instead of tokens.

> Muse Code supports 2 ways to pay for usage: - **Pay-as-you-go**: billed per token consumed with your Meta Model API key.
> — https://dev.meta.ai/docs/muse-code/subscriptions · Choose pay-as-you-go or a subscription section

**Free tier** · Yes — Self-serve throughout: install the CLI, sign in with a Meta account, then pay per token or subscribe from $5.00/month. No sales call, and no free Muse Code allowance is published.

> Meta Model API billing is usage-based: you're billed for the tokens your requests consume.
> — https://dev.meta.ai/docs/muse-code/auth · Billing and plans, Billing setup

**API access** · Yes — Beyond the CLI, a versioned session protocol plus a published TypeScript SDK lets your own program drive a Muse Code session; the underlying model is separately callable over the Model API.

> you can drive a Muse Code session from your own program over a stable, versioned session protocol. `muse serve` runs the protocol, and `muse schema` prints its JSON schema.
> — https://dev.meta.ai/docs/muse-code · Drive it from your own program section

**MCP server** · Partly — Muse Code is an MCP client — it connects stdio and streamable-HTTP servers, with OAuth sign-in — but the vendor documents no MCP server exposing Muse Code to another agent.

> Connect external tools through the Model Context Protocol (MCP). Declare servers in the `mcp_servers` block of your [settings file]
> — https://dev.meta.ai/docs/muse-code/extending · MCP servers section, opening paragraph

**Bring your own key** · Partly — You supply your own Meta Model API key, including in CI from a secret store, but inference still runs on Meta's hosted endpoint and no third-party key or own-cloud option is documented.

> For a non-interactive environment, or to avoid the browser flow, authenticate with an API key. Set it in the environment:
> — https://dev.meta.ai/docs/muse-code/auth · Use an API key section

## Buying it for a team

**A company can buy it** · Yes — Accounts are organised into teams with admins who hold payment methods and business information; pay-as-you-go billing is a team-level function, not an individual one.

> Only team admins can manage payment methods and business information.
> — https://dev.meta.ai/docs/muse-code/auth · Billing and plans, How to add payment section

**Seat model** · No — Nothing is priced per seat: billing is per token, per image or per audio minute, and the subscription is a flat monthly rate on one account.

> Meta Model API bills text models per token, image generation per image, and Muse Voice Transcribe per minute of audio processed, with no minimums or upfront commitment.
> — https://dev.meta.ai/docs/pricing-rate-limits · Pricing section, opening sentence

**Pooled budget** · Yes — Rate and token quotas are shared across the whole team rather than allocated per key, so multiple developers draw on one pool on pay-as-you-go.

> Limits apply **per team, not per API key**. If you use multiple keys in one team, all requests, tokens, images, and audio minutes count toward the relevant shared quota.
> — https://dev.meta.ai/docs/pricing-rate-limits · Rate limits section

**Admin controls** · Yes — Administrators can validate and apply managed defaults and policy documents and point the agent at a centrally administered hooks file, on top of per-session permission profiles.

> Enterprise administrators can validate a managed configuration document with `muse config validate --plane <defaults|policy> --file <path>`, or check the active managed configuration with `muse config status`.
> — https://dev.meta.ai/docs/muse-code/configuration · The settings file section

**Audit log** · Partly — Every run is an append-only event log covering model calls, tool calls and approval decisions, exportable as JSON — but it is a local session record, not a tenant-wide log of administrative action.

> Muse Code records every run as an append-only event log: model calls, tool calls and their results, and approval decisions.
> — https://dev.meta.ai/docs/muse-code/interactive · How sessions work section

**Single sign-on** · No — Exactly two documented sign-in paths: Meta browser sign-in or an API key. No SAML or OIDC, and Meta Managed Account users must use an API key.

> Muse Code needs a Meta credential before it can call the model. On first run it prompts you to choose how to authenticate: a browser sign-in or an API key.
> — https://dev.meta.ai/docs/muse-code/auth · Authentication and billing, opening paragraph

## What happens to your code

**Opt out of training** · Yes — Standard-tier models are contractually excluded from training, and the discounted contributor models that do train on your content are opt-in by model id and can be left behind by upgrading.

> Standard Services: Meta does not use your Content from Standard Services to train Meta Models.
> — https://dev.meta.ai/legal/commitments · Meta Model API Data Commitments, Standard Services paragraph

**Getting out** · Partly — Sessions export locally as transcripts or full JSON trajectories and account data can be requested from support within about five business days, but the export is a support ticket rather than self-serve.

> Your download will include: Account profile data (name, email, team membership). API usage history and conversation data. Billing records.
> — https://dev.meta.ai/help/accounts-and-login/download-account-data · How can I download my account information? answer

_Not established: Data residency, Certifications_

## Provenance

Every fact above comes from the vendor. Nothing here is independently corroborated yet.

## Coverage elsewhere

Pages about this product that the vendor does not control. The label states how independent each one is.

| Source | Kind | Read |
| --- | --- | --- |
| eesel.ai | Independent source | 2026-08-23 |
| wavect.io | Independent source | 2026-08-23 |
| venturebeat.com | Independent source | 2026-08-23 |

## What changed

- 2026-09-20 — Capabilities · Commercial terms · Free tier
- 2026-09-02 — Commercial terms · Free tier
- 2026-08-31 — Capabilities · Commercial terms · Free tier · Price
- 2026-08-23 — Summary

## Correcting this record

Anything marked "Not established" is a question we have not answered — it is not a No. If you represent Muse Code, or anything here is wrong or out of date, you can supply it with a source: the operations this site accepts from an agent are listed at https://theagentsindex.com/.well-known/agents.json — that catalogue needs no key to read, and each entry states its own auth.

Supplying evidence is free. Nothing that can be bought here changes what a field says, whether a gap is closed, or how this record is scored.