NEWOut of beta — now powered by Muse Spark 1.3

Meta’s first AI coding agent, in your terminal.

Muse Code is Meta’s AI terminal coding agent for your most complex coding workstreams. Build, debug, and ship: it plans the work, fans it out to persistent background agents in isolated git worktrees, and logs every decision.

$curl -fsSL https://dev.meta.ai/install.sh | bash
See it work ↓
muse — ~/payments-service
$
agent-1agent-2agent-3
In action

Watch it work.

Game development

Six features. One session. Zero collisions.

Muse Code built six game features simultaneously — each agent in its own isolated worktree. Nothing stepped on anything else, and every branch merged clean.

Web development

A one-line idea, shipped as a working app.

Describe the product in a sentence. Muse Code plans the build, stands up the stack, wires the pages, and validates the result — end to end, in your terminal.

Multimodal visual coding

From a video file to working software.

Drop an mp4 straight into the terminal — a fly-through of a vacation home becomes a finished marketing and booking page. Muse Spark reads video, images, audio, and PDFs, not just code.

How it works

From prompt to pull request in three steps.

01

Install in seconds

One command, then sign in with your Meta Developer account. No IDE, no plugins — just your terminal.

curl -fsSL https://dev.meta.ai/install.sh | bash
02

Describe the work

State the outcome in plain language. Muse plans the tasks and hands them to its persistent background agents, working in parallel.

$ muse "migrate auth to passkeys"
03

Review and ship

Audit the event log, replay any decision, then merge clean worktrees. You stay in control of every diff.

✓ 3/3 merged · all checks green
Why Muse Code

Accelerate your build.

Muse Code is a fast and accessible coding agent powered by Muse Spark.

Multi-agent by default

Multiple agents coordinate on every task — workers in parallel, reviewers in the background — so you ship faster without sacrificing quality.

agent-1refactor payments client
agent-2port calls to billing API
agent-3write integration tests

Isolated git worktrees

Every agent works in its own worktree — your working copy is never touched. No collisions, no clobbered files; changes merge only after review.

main— protected
├─../wt/agent-1✓ merged
├─../wt/agent-2● working
└─../wt/agent-3● testing

Fully auditable

Every action is transparent and traceable. Review generated code, understand AI decisions, and keep confidence with built-in visibility across your workflows.

12:04:01 agent-1 read src/payments/client.ts
12:04:03 agent-2 edit src/billing/api.ts +42 −17
12:04:05 agent-1 plan: split client into v2 adapter
12:04:07 agent-3 run pytest tests/integration 24 passed
12:04:09 main pause agent-2 → adjust plan
12:04:11 agent-2 resume with revised diff
12:04:14 main merge 3 worktrees clean
12:04:01 agent-1 read src/payments/client.ts
12:04:03 agent-2 edit src/billing/api.ts +42 −17
12:04:05 agent-1 plan: split client into v2 adapter
12:04:07 agent-3 run pytest tests/integration 24 passed
12:04:09 main pause agent-2 → adjust plan
12:04:11 agent-2 resume with revised diff
12:04:14 main merge 3 worktrees clean

Built for Muse Spark

Co-trained for tight integration: better tool use, fewer retries, higher-quality output than a generic wrapper — with a 1M-token context that holds your whole repo.

src/services/packages/tests/infra/docs/
1M
TOKEN CONTEXT
Muse Spark 1.3

The model and the harness, trained as one.

Muse Code runs on Meta’s Muse Spark 1.3, which replaced 1.2 as its model on September 2, 2026. The Spark models are co-trained with the Muse Code harness itself — rejection-sampled agent trajectories, goal conditioning, and context compaction — and trained extensively on long-horizon work. Version 1.3 finishes the same coding work with about 20% fewer tool calls and 25% fewer tokens than 1.2, and asks a clarifying question instead of guessing when it’s stuck.

Co-trained
model + tools, one system
1M tokens
repo-scale context
−25% tokens
vs. Spark 1.2, same work
Pricing

A monthly plan, or pay per token.

Meta sells it as a monthly plan from $5 or pay-as-you-go on its Model API. We just lay out the options.

CONTRIBUTOR
$0.10/ 1M input
$0.20 / 1M output · $0.002 / 1M cached input
Full agent capabilities, no caps
Radically low per-token pricing
Sessions may be used to improve Muse Code
Set up Contributor on Meta ↗
STANDARDPRIVATE
$1.25/ 1M input
$4.25 / 1M output · $0.15 / 1M cached input
Full agent capabilities, no caps
Standard API billing
Your data is never used for training
Get an API key from Meta ↗

Prefer a flat monthly fee?Everyday $5·High $15·Power $50Compare plans →

Questions, answered.

What is Muse Code?+

Muse Code is Meta’s AI coding agent for the terminal, built by Meta Superintelligence Labs. It launched in beta in August 2026, left beta on August 31, 2026, and now runs on the Muse Spark 1.3 model. It takes on complete software engineering tasks across large repositories — planning changes, writing code, and validating the results with persistent background agents.

Is Muse Code free?+

Installing it is free; using it is not. You pay with a monthly plan — Everyday Usage at $5, High Usage at $15, or Power Usage at $50 — or per token through the Meta Model API. Meta hasn’t announced a free tier. The cheapest option is the Contributor tier at $0.10 per 1M input tokens, which lets Meta use your sessions to improve its products.

Is Muse Code out of beta?+

Yes. Muse Code left beta on August 31, 2026, adding monthly plans, inter-session messaging, multi-agent Workflows, and an SDK in developer preview. Install it with the one-line command, sign in with your Meta developer account, then choose a plan or bind an API key.

Does Muse Code work on Windows?+

Not natively. Meta’s installer supports macOS and Linux. On Windows 10 or 11, install WSL2 (the Windows Subsystem for Linux), then run the same one-line installer inside your Linux shell and use Muse Code from there.

Is there a Muse Code desktop app or VS Code extension?+

No. Muse Code is a terminal program: Meta publishes no desktop app, IDE, or official VS Code extension. You can run it in VS Code’s integrated terminal. Downloads advertised as a “Muse Code desktop app” or “Muse Code for Windows” are not from Meta.

Is Muse Code the same as Meta Muse?+

No. Meta Muse, launched on September 8, 2026, is a consumer personal agent that handles errands like email, bookings, and forms, with a free tier plus Power ($20) and Maximum ($100) plans. Muse Code is the developer agent that writes code in your terminal, with its own $5, $15, and $50 plans. Both run on Meta’s Muse Spark models.

How do parallel agents avoid stepping on each other?+

Each background agent operates in its own isolated git worktree. Changes stay separate until the main agent reviews and merges them, so there are no file collisions or conflicting edits.

Is my code used to train models?+

Only if you choose the Contributor tier, where sessions may be used to improve the product in exchange for dramatically lower pricing. On the Standard tier, your data is never used for training.

How large a codebase can it handle?+

Muse Spark 1.3 supports a 1M-token context window — enough to plan, refactor, and verify across large monorepos end to end, not just single files.

Can I see what an agent did?+

Yes. Every tool call and decision is written to a full event log. You can watch live, pause or redirect an agent mid-task, and replay any run step by step afterward.

From the blog

Latest posts.

View all posts →