Muse Code

Meta's terminal AI coding agent that plans, writes, and validates code across large repositories using persistent bac...

Editorially reviewedChecked Sep 2026How we review

coding#write-code

Visit Website

API

Toolglade’s take

A capable, aggressively priced entry into the terminal coding-agent race, now upgraded to Meta's Muse Spark 1.3 model (September 2026) with more reliable long-instruction following and leaner, faster coding — though it's still a Mac-and-Linux beta whose cheapest tier asks you to trade your code and prompts for the discount.

About Muse Code

Meta's terminal AI coding agent that plans, writes, and validates code across large repositories using persistent background sub-agents — now running Meta's Muse Spark 1.3 model (September 2026).

Muse Code is Meta Superintelligence Labs' terminal coding agent, installed on macOS or Linux with a single shell command. It takes on repository-scale software tasks — planning changes, writing code, and validating results — coordinating a main agent loop with persistent async background sub-agents that stay active through a session to reduce latency and steering on multi-step work. A crash-safe event-log runtime records every model call, tool run, approval, and edit, so long-running jobs are replay-exact and can resume after a crash. It ships with bundled skills such as /plan, /grill, and /goal, and offers two token-priced tiers: a Standard pay-as-you-go tier and a much cheaper Contributor tier that trades a discount for permission to let Meta train on your prompts and completions. Muse Code launched in beta on August 5, 2026 on the Muse Spark 1.2 model. As of September 2026 it runs Meta's newer Muse Spark 1.3, which Meta says follows complex, long-form instructions more reliably, sustains longer-horizon and multi-tasking work, collaborates more actively (asking clarifying questions on ambiguous prompts and confirming before consequential actions), and improves adversarial and prompt-injection robustness — while coding with a cleaner, less verbose style that Meta engineers measured at roughly 20% fewer tool calls and 25% fewer tokens versus 1.2. Muse Spark 1.3 is also available through the Meta Model API. Muse Code competes with Claude Code, OpenAI's coding agent, Cursor, and Kiro.

Weighing your options?Compare Claude Code vs Muse Code (2026): Terminal Agents, Two Business Models, Cursor vs Muse Code (2026): AI-First IDE or Terminal Agent?, or see all Muse Code alternatives.

TL;DR

Muse Code is Meta Superintelligence Labs' terminal AI coding agent, launched in beta on August 5, 2026 and powered by the Muse Spark 1.2 model. It plans, writes, and validates code across large repositories, using persistent async background sub-agents and a crash-safe event-log runtime to sustain long-running tasks. Pricing is usage-based across two tiers: a Standard tier and a roughly 20x cheaper Contributor tier that lets Meta train on your data. Installable on macOS or Linux, it enters a crowded field alongside Claude Code, OpenAI's coding agent, Cursor, and Kiro.

Company overview

Muse Code is built by Meta Superintelligence Labs (MSL), the AI unit Meta stood up in June 2025 to pursue frontier models and personalized 'superintelligence,' led by Alexandr Wang. Muse Code and the Muse Spark 1.2 model that powers it were announced on August 5, 2026, part of a rapid cadence of releases from the lab.

As a division of Meta Platforms (NASDAQ: META), MSL is backed by one of the world's largest technology companies and its substantial AI compute and research resources, rather than by external venture funding.

Product features

Muse Code runs a main agent loop augmented by persistent async background sub-agents that remain active throughout a session, avoiding redundant context gathering and reducing latency on difficult, multi-step tasks. Its runtime writes every model call, tool run, approval, and edit to a local event log, making sessions replay-exact and restart-safe so long-horizon work can resume precisely after a crash.

It ships with bundled skills including /plan (turn a task into an approval-gated plan), /grill (stress-test that plan), and /goal (work toward a stated objective), and can ingest inputs such as reference files and even video. The underlying Muse Spark 1.2 model was co-trained with the Muse Code harness and tuned for long-horizon coding, from whole-repository generation to multi-hour GPU-kernel optimization.

Target market

Muse Code targets professional software engineers and AI/ML developers who work in the terminal on large, complex codebases and want an autonomous agent for planning, refactoring, debugging, and long-running tasks. Because it is currently a macOS and Linux beta, early adopters are likely to be individual developers and teams comfortable with command-line tooling and pay-as-you-go token pricing, including cost-sensitive users who may opt into the cheaper Contributor tier despite its data-sharing terms.

Buyer personas

End users

Software engineers, ML engineers, and platform developers who work primarily in the terminal on large repositories.

Buyers

Engineering leads, individual developers, and small teams managing their own tooling budgets under usage-based billing.

Key influencers

Developer-experience and AI-tooling advocates, staff engineers evaluating agentic coding tools, and open-source community voices.

Ideal customer profile

A terminal-native professional developer or team on macOS or Linux tackling repository-scale, long-horizon coding tasks and willing to work with a token-priced beta.

Funding & performance

Muse Code is a product of Meta Platforms, Inc. (NASDAQ: META), a publicly traded company, rather than a venture-funded startup. It is developed by Meta Superintelligence Labs and funded through Meta's corporate AI research and infrastructure budget. There is no separate startup funding round; the product's resources come from Meta's broader investment in foundation models and AI compute.

Pros & cons

Pros

Persistent async background sub-agents reduce latency and manual steering on multi-step tasks
Crash-safe event-log runtime resumes long-running jobs exactly where they stopped
Purpose-built Muse Spark 1.2 model co-trained with the coding harness
Standard tier is aggressively priced versus rival coding agents
Bundled /plan, /grill, and /goal skills structure and stress-test work
Backed by Meta's compute and available through the Meta Model API

Cons

Contributor tier's ~20x discount requires letting Meta train on your prompts and completions
Still an early beta, so stability and features may change
Limited to macOS and Linux; no Windows-native support at launch
No GUI, mobile app, or browser extension; terminal-only workflow
Contributor tier is rate-limited on a rolling five-hour window and available only in select countries
No free plan or free trial; billing is purely usage-based

Pricing plans

Standard

$1.25 / 1M input

~$4.25 / 1M output
cached input ~$0.15 / 1M
no data used for training

Contributor

$0.10 / 1M input

~$0.20 / 1M output
~20x cheaper
Meta trains on your prompts & completions

Key features

API

Multi-language

Integrations

Meta Model API, macOS terminal, Linux terminal, Git repositories

Input types

text, code

Output types

code, text

Best For

Repository-scale coding tasks, Long-horizon autonomous work, Terminal-based developers

Compare key features

View all alternatives →

Feature

Muse Code

Claude Code

Amp

Pricing

Paid

Free plan

Free trial

API

Yes

Team support

Yes

Frequently asked questions

What is Muse Code?+

Muse Code is a terminal-based AI coding agent from Meta Superintelligence Labs, launched in beta on August 5, 2026. It plans changes, writes code, and validates results across large repositories, using persistent async background sub-agents and a crash-safe runtime to handle long-running tasks.

Which model powers Muse Code?+

Muse Code launched on Muse Spark 1.2 and, as of September 2026, runs Meta's newer Muse Spark 1.3 — a coding- and agent-focused model tuned for long-horizon tasks. Meta reports 1.3 follows complex instructions more reliably and codes more efficiently (roughly 20% fewer tool calls and 25% fewer tokens than 1.2), with stronger prompt-injection robustness. Muse Spark 1.3 is also available through the Meta Model API.

What platforms does Muse Code support?+

Muse Code installs on macOS or Linux with a single shell command and runs in the terminal. There is no native Windows client, mobile app, or browser extension at launch.

How much does Muse Code cost?+

Pricing is usage-based. The Standard tier is reported at about $1.25 per million input tokens and $4.25 per million output tokens, with cached input around $0.15 per million. As beta figures, these may change.

What is the Contributor tier?+

The Contributor tier offers roughly 20x cheaper token pricing (reported around $0.10 per million input and $0.20 per million output) in exchange for permission for Meta to use your prompts and completions to train future models. It is rate-limited on a rolling five-hour window and available only in select countries.

How does Muse Code compare to Claude Code, Cursor, and Kiro?+

Like Claude Code, Muse Code is a terminal-centric agent focused on model quality and repo-wide autonomy, whereas Cursor is a GUI AI-native editor and Kiro emphasizes spec-driven, structured planning. Muse Code's differentiators are its persistent background sub-agents, crash-safe event log, and aggressive Standard-tier pricing, though as a beta it is less mature than established rivals.

What's new in Muse Spark 1.3?+

Released in Muse Code and the Meta Model API in early September 2026, Muse Spark 1.3 improves agentic and coding performance over 1.2: it sustains longer-horizon, multi-step work, collaborates more (asking clarifying questions and confirming before consequential actions), follows long-form instructions more reliably without dropping requirements, and codes with a cleaner, less verbose style (Meta measured ~20% fewer tool calls and ~25% fewer tokens). It also improves safety, including resistance to prompt injection, and Meta says a higher "max reasoning" mode and a Muse Spark open-weights release are on the roadmap.

Reviews

Loading reviews…

Compare

Compare Muse Code with other AI tools

Side-by-side pages for pricing, features, and best-fit use cases.

All comparisons →

45 views

Claude Code vs Muse Code (2026): Terminal Agents, Two Business Models

Both are terminal-native coding agents, but Claude Code ships on flat subscriptions while Muse Code is a beta with usage-based tokens and a data-for-discount tier.

Read comparison →

37 views

Cursor vs Muse Code (2026): AI-First IDE or Terminal Agent?

Cursor is a full AI IDE with subscription pricing; Muse Code is Meta's beta terminal agent billed per token. The choice is really editor versus command line.

Read comparison →

People also compare

Claude Code vs Muse Code (2026): Terminal Agents, Two Business Models45 views →Cursor vs Muse Code (2026): AI-First IDE or Terminal Agent?37 views →

From the blog

All articles →

Article

Meta Ships Muse Spark 1.3: A Leaner, Smarter Coding-Agent Model

Meta released Muse Spark 1.3, the model behind its Muse Code agent: more reliable on long tasks, ~20% fewer tool calls and ~25% fewer tokens, and stronger safety. Here's what changed.

Read →

Article

Muse Code Review (2026): Meta's Terminal AI Coding Agent

A hands-on look at Muse Code, Meta Superintelligence Labs' beta terminal coding agent powered by Muse Spark 1.2, its persistent background sub-agents, two-tier token pricing, and how it stacks up against Claude Code, Cursor, and Kiro.

Read →