Resident background agents
Background agents persist for a whole session rather than rebuilding context each turn — Meta's split between a chat loop and a working session.
Muse Code Cost Calculator
Estimate what a Muse Code session costs on each tier by entering the tokens a task actually consumes.
Standard
—
whole task · 40 turns
- per turn
- —
- 20 tasks / mo
- —
Contributor
—
whole task · same volume
- per turn
- —
- 20 tasks / mo
- —
Rates used: Standard $1.25 input, $0.15 cached input, $4.25 output per million tokens. Contributor $0.10 input, $0.002 cached input, $0.20 output per million tokens. Rates are as reported by third-party trackers and change during beta — check Meta’s pricing and rate limits documentation before relying on a number.
Install Muse Code in Three Steps
One command on macOS and Linux; the full install walkthrough has the detail.
Run the Muse Code install script
Run Meta's one-line installer:
curl -fsSL https://dev.meta.ai/install.sh | bashYou get a
musebinary in~/.local/bin; PATH fixes are in the Muse Code guide to installing.Sign in with your Meta account
Run
museonce; a device-style OAuth flow opens on auth.meta.com and waits on browser approval — which is why sign-in stalls in Docker. See the Muse Code guide to sign-in errors.Start Muse Code inside a repository
Run
musein a repository and describe the task; it answers with a plan you approve before anything runs. Windows needs WSL2 first, per the Muse Code guide to Windows.
What Muse Code Actually Does
What Meta describes in the launch post.
An append-only local event log
Every action lands in an append-only log on your machine, which Meta calls replay-exact and restart-safe: a crash mid-task keeps finished work.
Multimodal in a 1M-token window
Muse Spark 1.2 takes images, video and PDFs in a one-million-token context; the launch demo turned an mp4 house tour into a booking site.
Runs measured in hours
Meta reports a run optimising NVIDIA Hopper GPU kernels across 1,000+ tool calls and up to 24 hours; commenters noted the same study shows Muse losing to GPT-5.6 Sol.
Skills: /plan, /grill, /goal
/plan drafts a plan you approve, /grill attacks it, /goal drives to an outcome. Community reports suggest more, tracked in the Muse Code guide to skills and plugins.
No IDE, no app, no cloud runner
One terminal binary: no editor extension, no Mac app, no hosted runner, no roadmap for any. The narrowest surface in its category.
Muse Code Pricing: Standard and Contributor
No subscription, no free quota — the tier sets your rate and data terms.
| Tier | Input / M | Output / M | The trade-off |
|---|---|---|---|
| Standard | $1.25 | $4.25 | Meta states it does not train on your data |
| Contributor | $0.10 | $0.20 | You grant training rights; limited countries |
Rates are as reported, cross-checked 6 August 2026 against Meta's pricing page. Rate limits and cache rates: Muse Code guide to pricing. Per-task sums: cost calculator. Data terms: Muse Code guide to privacy.
Muse Code vs Claude Code, Codex, Cursor and Copilot
Muse Code bills per token, not per seat. Each row opens a full comparison.
| Comparison | Rival entry price | Pick the rival if |
|---|---|---|
| Muse Code vs Claude Code | $17–20 / month (as reported) | Windows or a desktop app |
| Muse Code vs Codex | $8–20 / month (as reported) | Cloud agents or a free tier |
| Muse Code vs Cursor | $20 / month (as reported) | An IDE, not a terminal |
| Muse Code vs GitHub Copilot | $10 / month (as reported) | Work that lives inside GitHub |
Rival prices are as reported for August 2026 and move often. Wider field: best AI coding agents and Muse Code alternatives; the model: Muse Spark 1.2.
What Developers Say About Muse Code
Two Hacker News reactions capture the loudest Muse Code themes this Muse Code guide keeps tracking: billing and trust.
What Changed Recently in Muse Code
A dated log this Muse Code guide refreshes weekly during beta.
- 2026-07-09 — Muse Spark 1.1 ships. The prior model: 1M-token context, multimodal, a month before Muse Code.
- 2026-08-05 — Muse Code beta and Muse Spark 1.2 launch together. Meta Superintelligence Labs publishes the launch post; the model is trained jointly with the harness and hits the Meta Model API that day.
- 2026-08-05 — third-party listings appear. OpenRouter lists it; Artificial Analysis measures independently: Intelligence Index 54 against 51 for 1.1, Terminal-Bench v2.1 at 80% against 78%, small SciCode and HLE regressions.
- 2026-08-06 — the community pushes back. The Hacker News launch thread reaches 278 points across 177 comments: no hard spending cap, Docker sign-in failures, selfie verification, Contributor regional limits, no Windows build.
- Still open. Muse Spark weights are unreleased; Mark Zuckerberg said more would be shared soon, as reported by The Register.
Every Friday this Muse Code guide re-reads the pricing page, the installer scripts and the model listings, then re-dates what changed. Where a number differs from your terminal, trust the terminal.
Muse Code Guide: Every Page in the Index
The complete Muse Code guide index, from install and pricing through comparisons, privacy and the model itself.
- What Is Muse Code?Where the Muse Code guide begins: the definition, the maker, the agent loop.
- Muse Code PricingBoth tiers, rate limits and the data terms attached to each rate.
- How to Install Muse CodeThe install walkthrough for macOS and Linux, sign-in and PATH fixes included.
- Muse Code Not WorkingSymptom-by-symptom fixes for command-not-found, 403, Docker and billing errors.
- Muse Code on WindowsWhy no native Windows build exists and how the WSL2 route works.
- Muse Code vs Claude CodeBenchmarks, platforms and token billing against a monthly seat.
- Muse Code vs CodexA terminal-only agent against a CLI, IDE and cloud stack with a free tier.
- Muse Code vs CursorTwo product categories: terminal agent versus full AI editor.
- Muse Code vs GitHub CopilotCredit-based agent sessions compared with straight per-token billing.
- Muse Code AlternativesWhere to go when the data terms or platform limits do not fit.
- Best AI Coding AgentsThe whole field ranked on price, platforms, models and agent quality.
- Muse Code PrivacyWhat each tier lets Meta do with your prompts and completions.
- Muse Spark 1.2Spec sheet and benchmark table for the model behind the agent.
- Muse Spark APIAPI access through the Meta Model API or OpenRouter, with rate limits.
- Muse Code Skills and PluginsThree documented skills plus the community-reported plugin surface.
- Muse Code Cost CalculatorToken spend per task, per day and per month on either tier.
- About This SiteWho writes this Muse Code guide and why it exists.
- MethodologyHow the Muse Code guide verifies claims and sources every number.
- Editorial PolicyCorrection, hedging and update rules for every Muse Code guide page.
- DisclaimerScope and limits of the information published here.
- Privacy PolicyWhat this website collects when you read it or use the calculator.
Frequently Asked Questions
Is Muse Code free?
No — no free tier. Meta bills per token and asks for a payment method up front. Standard is $1.25 input and $4.25 output per million tokens as reported; Contributor is $0.10 and $0.20 for training rights. Sums on the Muse Code cost calculator.
Does Muse Code include free Muse Spark 1.2 usage?
Nothing free is officially promised. Community reports suggest the Muse Spark 1.1 sign-up credit changed wording or went away, and no recurring quota is documented. Confirm on Meta's pricing page; updates in the Muse Code guide to pricing.
Can I set a spending limit on Muse Code?
Not a hard one. Billing sends email alerts but enforces no ceiling — the launch thread's loudest complaint. The workarounds are external: a virtual card with its own limit, or OpenRouter with a per-key cap. Both on the Muse Code cost calculator.
Does Muse Code work on Windows?
Not natively. The launcher recognises four targets — arm64 and x86 builds of macOS and Linux — so no Windows binary exists. WSL2 works, except sign-in's browser step must hand off to Windows. Sequence in the Muse Code guide to Windows.
Why does Muse Code return a 403 or hang at sign-in?
Two causes. A 403 printing "muse is available to Meta employees" is a leftover launcher gate: your account or region is not cleared for the tier. A Docker hang is structural — sign-in runs a device-style OAuth flow on auth.meta.com needing a terminal and a browser, and a container has neither. Fixes: the Muse Code guide to errors.
Does Meta train on the code you send to Muse Code?
Only on one tier. Standard: Meta states it does not train on your prompts or completions. Contributor: you grant training rights over both, and that grant buys roughly 12x cheaper input and 21x cheaper output. See the Muse Code guide to privacy.
How does Muse Code rank compared to Claude Code?
Below Opus 5, above GPT-5.6 Terra so far. Meta's self-reported Terminal-Bench 2.1 score of 82.9% trails Claude Opus 5 at 86.7% and beats GPT-5.6 Terra at 81.8%, all as reported second-hand; Artificial Analysis measured 80%. See Muse Code vs Claude Code.
Sources
- Meta AI Research — Introducing Muse Code and Muse Spark 1.2research.meta.ai
- Meta — Muse Spark pricing and rate limitsdev.meta.ai
- Meta for Developers — Muse Spark model pagedeveloper.meta.com
- Artificial Analysis — Muse Spark 1.2 measurementsartificialanalysis.ai
- OpenRouter — meta/muse-spark-1.2openrouter.ai
- Hacker News — Muse Code launch threadnews.ycombinator.com