Meta Muse Code
by Meta (Meta Superintelligence Labs)
Meta's terminal coding agent for large repositories, with persistent background agents and the most aggressive token pricing in the category
Muse Code is Meta's terminal-based AI coding agent, released in beta on August 5, 2026 and powered by the Muse Spark 1.2 model. It plans changes, writes code and validates results across large repositories, spawning parallel sub-agents in isolated git worktrees so the developer's main working copy is never touched.
Muse Code is a command-line coding agent from Meta Superintelligence Labs, released in beta on August 5, 2026 alongside the Muse Spark 1.2 model that powers it. It installs with a single command (curl -fsSL https://dev.meta.ai/install.sh | bash) and runs on macOS and Linux; there is no Windows build at beta. The design target is whole software-engineering tasks across large repositories - planning changes, writing code and validating results - rather than single-file completions. Two architectural choices set it apart. First, a set of async background agents stays alive for the entire session instead of being spawned per task, which Meta says cuts redundant information gathering and latency on long multi-step work; when a job is large enough, Muse Code fans out sub-agents that run simultaneously in isolated git worktrees, and Meta demonstrated building six features for a game in parallel with no collisions. Second, the runtime keeps a local append-only event log of every model call, tool run, approval and edit, making sessions replay-exact and restart-safe after a crash. Three skills ship by default: /plan converts a task into an approval-gated plan, /grill stress-tests that plan, and /goal drives toward completion. Muse Spark 1.2 is proprietary and closed-weight with a 1 million token context window, a departure from Meta's open-weight Llama posture. Pricing is the aggressive part: $1.25 per million input tokens and $4.25 per million output on the standard tier, or $0.10 and $0.20 on a contributor tier that requires permitting Meta to train on your prompts and completions.
Platform and developer-productivity leads at engineering organizations with large monorepos who already run terminal coding agents and are being squeezed by Claude Code or Codex token spend.
Parallel sub-agents in isolated worktrees close multi-file work across a big repository at roughly a third of frontier-agent token cost on the standard tier.
At a Glance
- Category
- Developer Tools
- Pricing
- Usage-based
- Target Market
- CTOs, VPs of Engineering, Platform Engineering Leads, Enterprise Developers, Developer Productivity Teams
- Deployment
- Cloud-only, API-based
- Founded
- 2004
- Headquarters
- Menlo Park, California, United States
- Team Size
- 500+
Key Features
- ✓Parallel sub-agents in isolated worktrees
Large jobs fan out across simultaneous sub-agents in separate git worktrees, so the main working copy is never touched.
- ✓Persistent async background agents
Agents stay alive across the whole session rather than respawning per task, cutting redundant context gathering and latency.
- ✓Replay-exact append-only event log
Every model call, tool run, approval and edit is logged locally, making a session restart-safe after a crash.
- ✓Bundled /plan, /grill and /goal skills
Turns a task into an approval-gated plan, stress-tests that plan for weaknesses, then drives it to completion.
- ✓One million token context window
Muse Spark 1.2 holds enough of a large repository in context to reason across files without constant re-retrieval.
- ✓Contributor pricing tier
Cuts input cost roughly 12x and output roughly 21x in exchange for explicit permission to train on your sessions.
- ✓Single-command terminal install
Installs via one curl command on macOS or Linux, with no IDE plugin or separate desktop application required.
Capabilities
Use Cases
- •Multi-feature parallel development
Meta demonstrated Muse Code building six features for a game simultaneously across isolated worktrees with no merge collisions.
- •Large-repository refactors and migrations
Background agents persist through long-horizon work spanning many files, planning changes and validating results without re-reading context each step.
- •Approval-gated change planning
The /plan and /grill skills produce a plan a human reviews and stress-tests before any code is written or committed.
- •Cost-constrained agentic coding at volume
Teams priced out of frontier coding agents can run the same workloads at $1.25 per million input tokens on the standard tier.
- •Unattended and CI-driven agent runs
The replay-exact event log makes long automated runs restartable after a crash instead of losing the whole session.
Ideal For
Best For
- ✓Large monorepos where a task spans many files and parallel sub-agents in isolated worktrees avoid merge collisions
- ✓Long-horizon refactors and migrations that benefit from background agents persisting across an entire session
- ✓Teams under pressure to cut coding-agent token spend, given standard-tier pricing well below frontier alternatives
- ✓Prototyping and internal experimentation where the 12x-cheaper contributor tier's data-sharing terms are acceptable
- ✓CI and unattended runs that need crash recovery, thanks to the replay-exact append-only event log
Not Ideal For
- ✗Regulated enterprises and any team with contractual code-confidentiality obligations, because the cheap contributor tier explicitly requires permitting Meta to train on prompts and completions
- ✗Windows-based development teams - the beta ships for macOS and Linux only
- ✗Buyers who need verified benchmark performance, since independent testing of Muse Spark 1.1 came in at 76.2 percent against Meta's published 80.0 and the 1.2 number is unverified
- ✗Organizations requiring self-hosted or air-gapped inference; Muse Spark 1.2 is closed-weight with no downloadable weights and the agent requires a login
- ✗High-throughput automation on the contributor tier, which is rate-limited to 60 requests per minute
Integrations
Deployment
Market Analysis
Pros
- ✓Standard-tier pricing at $1.25 per million input and $4.25 per million output undercuts frontier coding agents materially, which Meta AI chief Alexandr Wang framed as the product's central argument
- ✓Parallel sub-agents in isolated worktrees genuinely address the merge-collision problem in large repositories - Meta demonstrated six features built simultaneously without conflicts
- ✓The append-only local event log makes long sessions replay-exact and restart-safe, a real operational advantage for unattended runs
- ✓A 1 million token context window is competitive with the strongest agents in the category
- ✓Standard-tier data policy explicitly excludes prompts and completions from model training
Cons
- ✗Benchmark credibility is the loudest criticism: independent verification put Muse Spark 1.1 at 76.2 percent (±1.2) on Terminal-Bench 2.1 against Meta's published 80.0, a 3.8-point gap whose confidence interval is hard to explain as run variance - and the 82.9 percent claimed for 1.2 has no verified score at all
- ✗Hacker News commenters flagged within the hour that Meta benchmarked against Opus 5 rather than Fable 5 and GPT-5.6 Terra rather than Sol, and that Muse Spark 1.2 came out behind in the one kernel-optimisation case study where Sol does appear
- ✗The cheap contributor tier's default on-ramp sends developers' code and prompts into Meta for training, a trade most regulated or contractually bound enterprises cannot make
- ✗Muse Code requires a login and runs a closed harness against a closed-weight model, so HN developers noted there is no independent way to confirm what leaves the machine
- ✗Beta maturity shows - The Register's tester hit a wall where the agent reported no models visible and demanded payment to finish account setup - and there is no Windows build
- ✗Muse Spark is proprietary and closed-weight, a reversal of the open-weight approach Meta established with Llama
Pricing
Standard (pay-as-you-go)
From $1.25/1M input tokens
- ✓$1.25 per million input tokens
- ✓$4.25 per million output tokens
- ✓$0.15 per million cached input tokens
- ✓Prompts and completions are not used to train Meta's models
- ✓3,000 requests and 4M tokens per minute per team
Contributor (data-sharing)
From $0.10/1M input tokens
- ✓$0.10 per million input tokens
- ✓$0.20 per million output tokens
- ✓$0.002 per million cached input tokens
- ✓Requires explicit permission for Meta to train on prompts and completions
- ✓Rate-limited to 60 requests per minute
Metered purely on tokens with no seat licence. The standard tier is $1.25 per million input, $4.25 per million output and $0.15 per million cached input, and Meta commits that prompts and completions on this tier are not used for training. The contributor tier drops that to $0.10, $0.20 and $0.002 respectively - roughly 12x cheaper on input and 21x on output, among the cheapest in the coding-agent category - but only if you explicitly permit Meta to train on your sessions, and it is throttled to 60 requests per minute against the standard tier's 3,000 requests and 4 million tokens per minute per team. Muse Spark is also reachable through the Meta Model API and OpenRouter.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 6 on domains other than dev.meta.ai.
- 1.research.meta.ai — introducing muse code and muse spark 1 2
- 2.techcrunch.com — meta launches muse code an ai agent for large code bases
- 3.theregister.com — 5283717
- 4.marktechpost.com — meta superintelligence labs releases muse code
- 5.finance.biggo.com — 202608052250 Meta launches Muse Code AI coding agent
- 6.kingy.ai — muse code muse spark 1 2 benchmarks verified
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Niteshift
The full-stack cloud for coding agents — real environments, verified pull requests
Opik
Open-source tracing, evaluation and guardrails for LLM applications and AI agents
Latitude
Open-source observability and evaluation for AI agents — find and fix failures before production
Devin
The autonomous AI software engineer that plans, writes, tests, and ships production code inside your codebase