What Is Muse Code?

Muse Code is Meta's coding agent for the terminal and CI. You run muse inside a repository, and it plans a change, edits files, and runs commands, with approvals and an OS sandbox on by default. It runs on Muse Spark, Meta's own model family, which Meta says it co-trained with the agent harness. Meta released it in beta on August 5, 2026, and took it out of beta on August 31.

Two things set it apart from a single-agent assistant. It's multi-agent by default: persistent background agents run alongside the main agent, and a workflow engine can fan work out to parallel child agents. It also writes every model call, tool run, approval, and edit to a local append-only event log, which Meta says makes sessions restart-safe and replayable.

What it costs: a subscription at $5, $15, or $50 a month (Everyday, High, and Power Usage), or pay-as-you-go through Meta Model API at Muse Spark token rates. Pay-as-you-go has a cheap Contributor tier if you let Meta train on your prompts and completions.

What it runs on: Meta's pages disagree. The docs say macOS, Linux, and Windows. The product page says macOS and Windows. All three Meta blog posts that give install steps say macOS or Linux. The platform section below lays out each source.

Which model: Muse Spark 1.3 is available in Muse Code as of September 2, 2026, but Meta's docs still name muse-spark-1.2 as the default.

Timeline

How Muse Code works

The same muse binary runs two ways. muse opens an interactive terminal UI with slash commands, approvals, and live status. muse exec "<prompt>" runs one prompt to completion for scripts and CI, authenticated with a META_API_KEY environment variable.

Multi-agent by default

Meta describes a simple main agent loop plus a set of async background agents that stay alive for the whole session instead of being spawned per task. Meta says they carry out next steps and decide when to report back, which cuts repeated information gathering and the need for steering.

Workflows go further. You ask for one ("use a workflow…") or let Muse Code propose one, and it builds an orchestration, runs child agents in parallel or in stages, and returns one synthesized result. Per Meta's docs, a workflow can call up to 1,000 child tasks over its lifetime, with at most 16 active at once locally (the exact limit depends on your CPU). Child agents inherit the lead session's permission boundary and can't widen it. You watch and control runs in /workflows and can save a workflow to reuse later.

Event log, rewind, and session messaging

• Event log. Every model call, tool run, approval, and edit goes into one local append-only log. Meta says a session can resume exactly where it stopped after a crash. Much of the recent changelog is resume and recovery fixes, which fits that design.
• Rewind. /rewind or a double Esc branches the conversation from an earlier message and keeps the original session. Rewind changes conversation history only. Meta's rewind docs say it "does not restore, revert, or delete workspace files," so you undo code changes with Git. Meta's plans post says you "confirm before any work is undone," which reads broader than what the docs describe.
• Session messaging. Separate live Muse Code sessions on the same machine can send each other handoffs, review requests, and status updates of up to 8,192 bytes over a local Unix socket. A message is treated as unverified data. It can't approve a tool call or change permissions. It works only on macOS and Linux, and only for interactive sessions.
• SDK (developer preview). muse serve exposes a versioned session protocol (Muse Session Protocol, or MSP) over standard I/O. Meta publishes @muse-code/sdk for TypeScript and a Python package, muse-code-sdk. The official repository is meta-models/muse-code-sdk on GitHub (MIT license). It holds the SDK and docs, not the Muse Code agent itself.

Where approval gates, sandboxes, and recovery logs fit in an agent architecture

concept↗

Platform support

This is the most inconsistent part of Meta's material. Here's what each official source said when we read it on September 26, 2026:

Models

Muse Code runs Muse Spark. Meta lists four Muse Spark model IDs you can pick in Muse Code. Each version comes in a Standard and a Contributor variant, and the two differ in price and data use:

Default model conflict. Meta's Muse Code overview and configuration docs both say "The default model is muse-spark-1.2." Meta's Muse Spark 1.3 post says 1.3 "is now available on Muse Code," and the product page lists 1.3 first. The unofficial musecodes.io says 1.3 "replaced 1.2" in Muse Code on September 2. No Meta page we read says 1.3 is now the default. You can pick a model with --model or /models, so if you want 1.3, choose it explicitly.

Reasoning effort goes up to xhigh and max. The configuration docs say Muse Code offers max on both Standard and Contributor tiers. ultra is a client-side setting that maps to the highest tier the provider supports and makes Muse Code delegate more aggressively.

Pricing, access, and current limits

Subscription plans

Prices are from Meta's product page. Plan contents and the 5x and 20x multipliers match on the product page and in the subscriptions docs, read September 26, 2026. Meta doesn't define a "prompt" or explain why the Everyday range runs from 10 to 50, so you can't turn these plans into a number of tasks. We haven't measured it either.

Pay-as-you-go through Meta Model API

Without a subscription, Muse Code bills your Meta Model API key per token:

Meta describes Contributor pricing as "heavily discounted" in exchange for permission to use your prompts and completions to train future Meta models. Standard-tier prompts and completions aren't used for training. For a coding agent, "prompts and completions" can include your source code, so pick Contributor only for code you're allowed to share. Meta's subscriptions page says how it uses your inputs and outputs "depends on the models you select." It doesn't say which tier subscription usage runs on.

Meta doesn't list a free Muse Code plan on any page we read.

Meta's benchmark claims

These are Meta's own results, not independent measurements. We haven't reproduced them.

Don't combine numbers across these two pages. The 1.3 page lists Muse Spark 1.2 (xhigh) at 55.0 on DeepSWE v1.1, while the launch chart gave 1.2 59.3%. Opus 5 (max) appears at 74.0 on one page and 65.0% on the other. Meta doesn't explain the difference on either page, so the setups probably differ. Each chart is only useful for comparisons within that chart.

Model scores also aren't agent scores. The same model can behave differently in a different harness, and none of these benchmarks is your repository.

Where Muse Code might fit

How to run any coding agent with scoped tasks, validation gates, and review before trusting its diffs

workflow↗

Muse Code vs Meta Muse, Muse Spark, and Meta Model API

Meta's docs sum up the Model API relationship: Muse Code and the Model API are "two ways to use the same model."

Meta Muse, Meta's consumer agent, and how its plans differ

entity↗

Choosing a model by role before choosing an agent that wraps it

concept↗

What we haven't verified

• We haven't run Muse Code on any platform. On September 26, 2026, both official installer URLs (install.sh and install.ps1) returned HTTP 200. We didn't run either.
• We haven't measured how many tasks fit in a subscription, what a task costs pay-as-you-go, or how fast the agent is.
• Which model new sessions default to in the current build (the docs say 1.2).
• Whether workflows ship in Linux, Intel macOS, or Windows builds.
• Which data-use tier subscription usage falls under.
• Whether the subscription credential works with muse exec and muse serve, or only with the interactive CLI. The docs say "only through the Muse Code CLI."

How we label vendor documentation, measurements, inference, and community reports

concept↗

Change log

Sources & Documentation

1. Meta: Muse Code product page — Plans ($5, $15, $50), 5x and 20x multipliers, model and price table, "Available for MacOS and Windows."
2. Meta docs: Muse Code overview — Terminal and CI agent, installers for macOS, Linux, and Windows, default model muse-spark-1.2, Windows differences, SDK.
3. Meta docs: Muse Code changelog — Latest entry 1.2.1, including Auto-review default, /rewind, and Claude Code/Codex migration skill. No release dates.
4. Meta docs: Muse Code subscriptions — Plan contents, CLI-only subscription key, pay-as-you-go for other keys, limits, upgrades, and cancellation.
5. Meta docs: Authentication and billing — Browser sign-in or API key, META_API_KEY for CI, usage-based billing.
6. Meta docs: Configuration and context — Default model muse-spark-1.2, reasoning effort levels, MEMORY.md prompt-injection warning.
7. Meta docs: Permissions and safety — Approval and sandbox layers, permission profiles, Auto-review default.
8. Meta docs: Run multi-agent workflows — Workflow limits and the missing workflow engine in the public Apple silicon macOS package.
9. Meta docs: Rewind a conversation — Rewind changes conversation history only, not workspace files.
10. Meta docs: Coordinate sessions with messages — Local-only messaging, 8,192-byte limit, macOS and Linux only.
11. Meta docs: Pricing and rate limits (Meta Model API) — Standard and Contributor pricing, training-data tradeoff, per-team rate limits.
12. Meta research blog: Introducing Muse Code and Muse Spark 1.2 (Aug 5, 2026) — Beta launch, background agents, event log, bundled skills, co-training, Meta's benchmark charts.
13. Meta developer blog: Muse Code: New plans and features (Aug 31, 2026) — Out of beta, subscriptions, inter-session messaging, workflows, rewind, SDK preview, "macOS or Linux."
14. Meta research blog: Introducing Muse Spark 1.3 (Sep 2, 2026) — 1.3 available in Muse Code, Meta's tool-call and token efficiency claim.
15. Meta: Muse Spark 1.3 model page — Relationship only. Meta's 1.3 benchmark table and pricing.
16. Meta: Meta Model API product page — Relationship only. Public preview, OpenAI-SDK-compatible access to Muse Spark.
17. GitHub: meta-models/muse-code-sdk — Official SDK and protocol docs (MIT license, developer preview).
18. musecodes.io (unofficial, not affiliated with Meta) — Cited only for its conflicting platform and default-model claims.