Muse Code is a terminal-resident coding agent from Meta, released as an early beta on August 5, 2026. It handles the full cycle across an entire large repository: planning, code generation, verification, and debugging. When a task grows large, multiple sub-agents work in parallel on independent Git worktrees while a reviewer checks their results in the background. It runs on a dedicated model, “Muse Spark 1.2” (context window up to 1 million tokens), and supports integration with MCP servers as well as headless execution in CI environments.
Key Features
- Multi-agent architecture: As a task grows, several persistent sub-agents work in parallel on independent Git worktrees. While the workers make progress, a reviewer checks their output in the background
- Long context via Muse Spark 1.2: Powered by “Muse Spark 1.2,” a model tuned specifically for Muse Code, with a context window of up to 1 million tokens. The model is positioned to improve code generation, complex debugging, codebase comprehension, and end-to-end development workflows
- Resumability through a local event log: API calls, tool executions, developer approvals, and file changes are all written to a local event log before they run. If the process crashes mid-task, you can resume from the point where it stopped
- Extensibility as an MCP client: Declare external MCP servers in the
mcp_serversblock of the config file and use them as tools (Muse Code itself is an MCP client, not a server). Two transports are supported,stdioandstreamable_http, and each server can be markedrequired(abort on failure) oroptional(skip on failure). Note that MCP tools run as ordinary child processes outside the sandbox, so you are expected to connect only servers you trust - Headless execution in CI:
muse exec "<prompt>"runs a single prompt without any interactive session. In non-interactive environments you need to set theMETA_API_KEYenvironment variable, which is designed with CI pipeline integration in mind
Pricing
Muse Code has no subscription plans; it is offered purely on usage-based, per-token billing.
| Tier | Price | Highlights |
|---|---|---|
| Standard | Input $1.25 / 1M tokens, output $4.25 / 1M tokens (cached input $0.15) | Prompts are not used to train Meta’s models |
| Contributor | Input $0.10 / 1M tokens, output $0.20 / 1M tokens (cached input $0.002) | Significantly cheaper than Standard in exchange for agreeing to let Meta use your prompts and generated output to train its AI models |
Pricing is current as of August 2026. There is no free tier, and both tiers require a registered payment method before you can start. Check the official documentation for the latest pricing and terms.
Pros and Cons
✅ Pros
- Multiple sub-agents running on independent Git worktrees let you push several tasks forward in parallel, even in a large repository
- Muse Spark 1.2’s 1-million-token context window makes it easier to work with an entire large codebase in view
- The local event log lets long-running tasks resume from where they stopped if they crash partway through
- Support for MCP server integration and headless execution in CI makes it easy to fit into existing development workflows
- The Standard tier is priced below alternatives such as Anthropic’s Sonnet 5 series ($2 / 1M input, $10 / 1M output)
⚠️ Cons
- No subscription plan or free tier; getting started requires a registered payment method and usage-based charges
- Choosing the cheaper Contributor tier means your prompts and generated output are provided to Meta for AI model training
- MCP tools run outside the sandbox, so connecting an untrusted MCP server carries risk
- Only Mac and Linux are supported, and as of August 2026 the product is still an early beta
Comparison with Similar Services
| Criteria | Muse Code | Claude Code | OpenAI Codex CLI | Gemini CLI |
|---|---|---|---|---|
| Provider | Meta | Anthropic | OpenAI | |
| Strength | Parallel sub-agent work on independent Git worktrees | Complex multi-file refactoring | Interactive, intent-driven task handling | Google product integration, large-context handling |
| Billing model | No subscription, per-token usage billing only | Included in paid Claude plans (Pro and above) | Included in ChatGPT plans, API key usage billing also available | Free tier available, paid plans optional |
| Underlying model | Muse Spark 1.2 (up to 1M tokens) | Claude (large context depending on plan) | GPT series | Gemini (supports large context) |
| Supported OS | Mac/Linux | Mac/Linux/Windows | Mac/Linux/Windows | Mac/Linux/Windows |
Who Is This For?
- Developers who want to hand an entire large repository to an agent, from planning through code generation and verification
- Teams that want to run multiple feature additions and bug fixes in parallel as independent units of work
- Anyone who wants to integrate an agent into an existing workflow through MCP or headless execution in CI
- Users who prefer paying per token for what they use rather than a monthly subscription
Summary
Muse Code is Meta’s terminal-resident coding agent, built to take on an entire large repository end to end, from planning through code generation, verification, and debugging. Its strengths are parallel processing by multiple sub-agents on independent Git worktrees with a background reviewer, and a long 1-million-token context. On the other hand, there is no subscription or free tier, and if you choose the cheaper Contributor tier, keep in mind that your prompts will be used for Meta’s AI training. As of August 2026 it is still an early beta, so it is worth watching for upcoming feature expansions and changes to the pricing structure.