Auditable repository work
Teams that want model calls, approvals, tool runs and edits recorded in an append-only event stream for review or replay.
Independent tool overview
Muse Code at a glance
Muse Code is Meta's beta terminal coding agent for macOS and Linux, powered by the hosted Muse Spark 1.2 model. It can inspect a repository, propose and edit code, run tools, validate results and coordinate persistent background agents. Its append-only event log and official fail-closed sandbox recipe make it unusually strong for auditable, long-running work, but the log does not make inference local: repository context still goes to a hosted model. Muse Code has no separate seat fee; usage is token-billed through Meta's Model API. Standard processing costs more and is labeled as not used to improve Meta products, while the heavily discounted Contributor tier permits submitted data to be used for improvement. Keep approval gates and sandbox enforcement on, isolate concurrent work, protect the event log, and never give an autonomous beta agent unsupervised production authority.
Visit the official Muse Code site ↗
- Product type
- Terminal AI coding agent
- Developer
- Meta
- Status
- Beta
- Primary model
- Muse Spark 1.2
- Platforms
- macOS and Linux
- Billing
- Usage-based Model API tokens
- Agent model
- Foreground agent plus persistent background agents
- Auditability
- Append-only local event log and replay
- Sandbox
- Available; can also be disabled
- Open weights
- Not announced for Muse Spark 1.2 as of review
- Reviewed
- August 31, 2026
Overview
What Muse Code is
Muse Code is Meta's terminal-based coding agent, launched in beta in August 2026 and co-developed with the Muse Spark 1.2 model.
It is currently documented for macOS and Linux. Meta does not document a native Windows installation path on the launch page.
The agent can plan changes, edit files, run tools and validate work across large repositories. Multiple persistent background agents can continue working asynchronously and decide when to report back.
Muse Code stores an append-only local event log covering model calls, tool runs, approvals and edits. Meta says this supports exact replay, restart recovery and auditable CI workflows.
The launch version bundled approval-oriented skills including planning and plan review. Meta's current cookbook also demonstrates goal tracking, acceptance testing, adversarial review and steerable agent teams.
Meta's official contained-execution recipe uses an operating-system sandbox with limited write locations and restricted network access. It runs a live enforcement probe and refuses to continue if the sandbox is not actually working.
The agent can be run with sandboxing disabled, including through a --yolo option documented in Meta's cookbook. Do not use that mode on untrusted code, production systems or a workstation containing accessible secrets.
Muse Spark 1.2 is a hosted model available through Muse Code, Meta's Model API and provider integrations. Meta had not announced open weights for Spark 1.2 by this review; Muse Glimmer's open release is a different model.
Meta published a kernel-engineering case that ran for more than 1,000 tool calls and up to 24 hours. That is a vendor-selected case study, not a typical completion-time or reliability guarantee.
The local event log does not mean local inference. Prompts, code and other context required by the task are sent to the selected hosted model service.
Pricing is usage-based rather than a monthly Muse Code subscription. The Standard and Contributor endpoints have materially different data-use terms, so the cheapest rate is not suitable for every repository.
Muse Code is beta software. Models, prices, plan labels, skills, limits and installation behavior can change, so teams should verify the signed-in developer portal before adoption.
Use cases
Who Muse Code is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Long-running engineering tasks
Multi-step work that benefits from persistent background agents, explicit acceptance criteria and restart-safe state.
Large-codebase investigation
Exploring an unfamiliar repository, tracing behavior and turning findings into an approval-gated implementation plan.
Parallel experiments
Trying independent approaches in isolated worktrees before a human chooses which changes to merge.
Meta Model API users
Developers already comfortable managing Meta credentials, usage billing and data-use choices.
Capabilities
Core Muse Code features
Persistent background agents
Runs asynchronous agents that can remain active through a session and report when their assigned work reaches a useful checkpoint.
Append-only event log
Records model calls, tool execution, approvals and edits for recovery, audit and deterministic replay workflows.
Planning and review skills
Supports approval-gated plans, adversarial plan review, documented research and task-specific skill workflows.
Goal tracking
Meta's cookbook demonstrates pinned objectives, progress state, step probes and a completion audit that runs acceptance tests.
Contained execution
An official recipe restricts writes and network access, tests enforcement live and fails closed when isolation is not real.
Large-repository context
Muse Spark 1.2 is designed for code generation, debugging, codebase understanding and long-horizon whole-repository work.
Multimodal development
Muse Spark 1.2 can use visual and video inputs for workflows such as turning a reference recording into a site or artifact.
Model API billing choice
Users can choose Standard processing or a lower-cost Contributor endpoint with different data-use terms.
Process
How the Muse Code workflow works
Step 1
Classify the repository
Decide whether the code is public, synthetic, internal or regulated before choosing an API endpoint and submitting any context.
Step 2
Review installation behavior
Inspect Meta's installer and use its no-profile-modification option when appropriate; the default installer can add Muse to shell startup PATH configuration.
Step 3
Start from a clean worktree
Record the current branch and uncommitted files, then isolate agent work so user changes and parallel agents cannot overwrite one another.
Step 4
Protect secrets
Remove accessible credentials, environment files, customer data and production tokens from the agent's working scope.
Step 5
Enforce containment
Use the official sandbox or an equivalent least-privilege environment with only required filesystem and network access.
Step 6
Define acceptance criteria
State the intended behavior, tests, non-goals, performance limits and forbidden actions before implementation begins.
Step 7
Approve the plan
Have Muse Code inspect the repository and propose a bounded plan; review the affected files and risks before allowing edits.
Step 8
Separate parallel work
Give background agents independent tasks and worktrees, then review their outputs before combining changes.
Step 9
Inspect every diff
Review generated code, dependency changes, migrations and configuration manually; do not treat passing AI-written tests as independent proof.
Step 10
Run independent validation
Execute the repository's tests, type checks, linting, build and relevant security checks outside the model's own completion narrative.
Step 11
Protect the audit trail
Treat event logs as sensitive because they can contain prompts, source code, commands, outputs and accidentally exposed secrets.
Step 12
Keep deployment human-controlled
Require explicit approval for releases, database migrations, credential changes, destructive commands and billing-impacting actions.
Cost
Muse Code pricing and free plan
Muse Code is currently a beta agent without a separate seat subscription. Usage is billed through the Muse Spark 1.2 Model API endpoint. Meta's Standard rate is $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens, and is labeled as not used to improve Meta products. The Contributor endpoint is much cheaper at $0.10 input, $0.002 cached input and $0.20 output per million tokens, but permits submitted data to be used to improve Meta products. No durable free quota is documented on the public beta launch page. Rates and data-use labels were checked August 31, 2026; verify the signed-in developer portal before sending code.
Muse Code beta
No separate seat fee
Install the agent and pay for the Muse Spark 1.2 tokens it consumes.
- Requires a compatible Meta developer account and API access
- macOS and Linux are documented
- Long-running and parallel agents can materially increase token usage
Muse Spark 1.2 Standard
$1.25 input / $0.15 cached / $4.25 output per 1M tokens
The higher-cost endpoint labeled as not using submitted data to improve Meta products.
- Usage-based billing
- Standard endpoint: muse-spark-1.2
- Review retention, access control and contract terms for confidential code
Muse Spark 1.2 Contributor
$0.10 input / $0.002 cached / $0.20 output per 1M tokens
A deeply discounted endpoint whose submitted data can be used to improve Meta products.
- Usage-based billing
- Contributor endpoint: muse-spark-1.2-contributor
- Use only with public, synthetic or explicitly approved code and prompts
Pricing checked . Check current pricing at the source ↗
Assessment
Muse Code strengths and limitations
Where it stands out
- Persistent background agents support long-running and parallel engineering work.
- The append-only event stream provides stronger audit and recovery primitives than a simple chat transcript.
- Official recipes show fail-closed sandbox enforcement rather than relying only on model instructions.
- Approval-oriented planning and review skills encourage deliberate work before edits.
- Goal tracking and completion audits can keep multi-step tasks aligned with explicit acceptance criteria.
- Muse Spark 1.2 is designed for whole-repository reasoning, debugging and end-to-end implementation.
- Usage billing avoids a separate per-seat Muse Code subscription.
- Two endpoints let eligible users trade price against data-use permission.
- Meta publishes practical cookbook examples for replay, containment and coordinated agents.
What to consider
- Muse Code and Muse Spark 1.2 are beta products whose behavior, pricing and availability can change.
- Only macOS and Linux are documented; no native Windows setup is presented on the launch page.
- The official one-line installer executes a downloaded script and may modify shell startup files unless configured otherwise.
- A local event log does not make the model local; necessary repository context is still sent to a hosted service.
- Standard's not-used-for-improvement label does not by itself establish a particular compliance, retention or security posture.
- Contributor pricing permits submitted data to be used for product improvement and is inappropriate for confidential code without explicit organizational approval.
- Persistent subagents can create conflicting edits, larger review burdens and unexpectedly high token spend.
- The sandbox can be disabled, removing a material protection against unsafe or destructive tool use.
- Event logs can themselves become sensitive records containing code, commands, outputs and secrets.
- AI-generated tests may repeat the same misunderstanding as the generated implementation.
- Vendor case studies and evaluations do not guarantee reliability on a team's own repository.
- Muse Spark 1.2 weights were not publicly released as of this review, creating dependence on hosted access.
- No durable public free allowance is documented, so experimentation can incur usage charges.
- An autonomous agent should not receive unsupervised authority over production deployments, migrations, secrets or destructive operations.
Compare
Muse Code alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Codex
A strong alternative for agentic terminal and cloud coding workflows with repository-aware planning, implementation and review.
Explore Codex →Coding
Claude Code
A mature terminal coding agent with strong codebase reasoning, extensibility and permission controls.
Explore Claude Code →Coding
Cursor
Better for developers who prefer an AI-native graphical editor with inline changes and interactive repository navigation.
Explore Cursor →Coding
Gemini CLI
An open-source terminal alternative for users invested in Google's Gemini models and tooling ecosystem.
Explore Gemini CLI →Questions
Muse Code FAQs
What is Muse Code?
Muse Code is Meta's beta terminal coding agent. It uses Muse Spark 1.2 to inspect repositories, plan and edit code, run tools, validate results and coordinate persistent background agents.
Is Muse Code free?
The agent has no separate documented seat fee, but Muse Spark 1.2 usage is token-billed through Meta's Model API. No durable public free quota is documented on the launch page.
How much does Muse Code cost?
As checked August 31, 2026, Standard costs $1.25 input, $0.15 cached input and $4.25 output per million tokens. Contributor costs $0.10, $0.002 and $0.20 respectively but permits submitted data to be used to improve Meta products.
Can I use the Contributor tier with private code?
Only if your organization explicitly approves the data-use terms and has the right to submit every relevant code, prompt and file. Use Standard or another approved environment for confidential repositories.
Does Muse Code run locally?
The CLI and event log run locally, but Muse Spark 1.2 is a hosted model. Repository context required for a task is sent to the configured model service.
Is Muse Spark 1.2 open source?
Meta had not announced open weights for Muse Spark 1.2 by August 31, 2026. Do not confuse it with other Muse models that Meta has released openly.
What operating systems support Muse Code?
Meta's launch instructions document macOS and Linux. A native Windows installation path is not documented there.
Is Muse Code safe to run?
It can be run more safely with a clean worktree, least-privilege credentials, sandbox enforcement, restricted network access and human approval. Disabling containment or exposing production secrets materially increases risk.
What does Muse Code's event log record?
Meta says the append-only log records model calls, tool runs, approvals and edits. Protect it like source code because it may also contain sensitive prompts, outputs or secrets.
Can Muse Code work autonomously for hours?
It supports persistent long-running agents, and Meta published an extended vendor case study. That does not justify unsupervised production access; use checkpoints, budgets, acceptance tests and human review.
Should I use --yolo mode?
Not on untrusted code, production systems or a machine with accessible secrets. That option disables the sandbox protection described in Meta's official cookbook.
Does Muse Code replace code review?
No. A human should inspect the diff, dependencies and migration risk, then run independent tests, builds, security checks and deployment review.
Bottom line
Our Muse Code verdict
Muse Code is one of the more technically interesting terminal agents because Meta pairs persistent background work with an append-only event stream, replayable sessions and a practical fail-closed containment recipe. It is a credible fit for experienced teams that want auditable, long-horizon repository automation and can manage token budgets, isolated worktrees and strict approval boundaries. It is not a low-governance shortcut: the product is beta, the model is hosted, background agents amplify both throughput and risk, and the discounted Contributor endpoint carries consequential data-use terms. Start with a non-sensitive repository, Standard processing, sandbox enforcement and a human-controlled merge and deployment path.
Visit Muse Code website ↗Similar Coding Tools
View all Coding AI tools →
MAI-Cyber-1-Flash
MAI-Cyber-1-Flash - Microsoft's new cybersecurity model for securing codebases
Origin
Origin - Cursor's code hosting with GitHub sync and built-in agents
Cursor Router
Cursor Router - Cursor's auto-router picking the cheapest capable model per task
Slack Code
Slack's new channels for building software with AI agents as a team
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.