What is Muse Code?
Muse Code is a command-line coding agent announced on 5 August 2026 in Meta's AI research blog and promoted by Meta CEO Mark Zuckerberg. Unlike Claude Code from Anthropic or OpenAI's ChatGPT Codex, it has no dedicated desktop app — it runs purely in the terminal on macOS and Linux and installs with a single command:
curl -fsS https://dev.meta.ai/install.sh | bash
Meta positions it for complex software engineering tasks: planning, implementation and validation of changes across large codebases. The company says it is "our next step toward the frontier", with larger and more capable models already on the roadmap.
Parallel sub-agents in isolated worktrees
The core idea behind Muse Code is a set of asynchronous background agents that stay active for the entire session instead of being re-spawned for every task. They build up context over time, carry out next steps on their own and decide when to report back to the main agent — reducing latency and the need for manual steering on difficult, multi-step jobs.
When a task is big enough, Muse Code fans out to separate sub-agents working in parallel in isolated worktrees. The developer's working copy is never touched. Meta says that in testing, the tool built six features for a game simultaneously with no collisions — exactly the scenario where single-threaded agents stall.
Crash-proof by design: the local event log
Every model call, tool run, approval and edit is appended to a local event log before it executes. This single source of truth makes the runtime replay-exact and restart-safe: if the agent crashes mid-task, it picks up precisely where it stopped — no lost work and no re-prompting. That design allows Muse Code to take on long-running jobs that most agents give up on.
Muse Code also ships with built-in skills: /plan turns a task into an approval-gated plan, /grill stress-tests that plan until it holds up, and /goal works toward completion of a stated objective.
Muse Spark 1.2: the model behind the agent
Muse Code is powered by Muse Spark 1.2, a coding-focused update of the Muse Spark 1.1 model released in July 2026. Meta significantly scaled up training compute on coding tasks and expanded the diversity of training environments. Crucially, the model was co-trained together with the Muse Code harness, including rejection-sampled tool trajectories and recipe optimisations for sub-agents and context compaction — so it performs best when paired with Meta's own tooling.
The training also used a self-improvement loop: Muse Spark 1.1 generated challenging coding environments and instruction templates, then graded candidate solutions against them, producing a scalable dataset for 1.2. The result, according to Meta, is better code generation, complex debugging, codebase understanding and long-horizon work such as whole-repository generation.
Case study: 1,000+ tool calls on NVIDIA Hopper
As a demonstration, Meta let Muse Spark 1.2 optimise GPU kernels — KDA and MLA kernels for NVIDIA Hopper — over more than 1,000 tool calls and up to 24 hours of runtime. The agent wrote, compiled, profiled and iteratively improved performance against a FLA Triton baseline, combining a chunk-parallel preparation kernel with a sequential inter-chunk scan and KDA-specific optimisations. Meta reports substantial gains that continued well beyond the initial exploration phase.
Pricing: undercutting Claude Code and Codex
Pay-as-you-go access to Muse Spark 1.2 costs $1.25 per million input tokens and $4.25 per million output tokens — approximately €1.15 and €3.90 at an exchange rate of roughly 0.92 EUR/USD. That undercuts comparable frontier coding models, according to AI Expert. A contributor tier for developers willing to help improve the model costs more than ten times less than the standard rate, making it nearly free to try. The model supports a context window of up to 1 million tokens and is available both inside Muse Code and through the Meta Model API with expanded global access.
What it means for European developers
For European teams, Muse Code has several concrete attractions. It runs on Linux, so it slots into existing CI/CD and DevOps workflows without a new application. The contributor tier removes the cost barrier for evaluation, and the Meta Model API is accessible from the EU. The local event log is also a privacy-friendly detail: every record of what the agent did stays on the developer's own machine rather than in a cloud dashboard — a meaningful point for companies dealing with GDPR and data-residency requirements.
The launch extends Meta's push into agentic coding that began when safety harnesses for AI coding agents became a hot topic in the open-source community. With Llama replaced by the Muse Spark family in April and a paid API introduced in July, Meta is signalling that it plans to compete on both capability and price — and the coding agent price war with Claude Code and Codex has only just begun.
