What Is Muse Code?
Meta's terminal coding agent, released in beta in August 2026, plans changes, writes code, and validates results across large repositories, and can reach business data through MCP servers.
Muse Code is Meta's AI coding agent that runs in the terminal, released in beta on August 5, 2026. Powered by the Muse Spark 1.2 model, it takes on complete software engineering tasks across large repositories, planning changes, writing code, and validating the results, and it can fan work out to parallel sub-agents in isolated git worktrees. It supports the Model Context Protocol (MCP), so it can also reach external tools like email, calendar, and drive.
Where Muse Code came from
Meta announced Muse Code on August 5, 2026 as its first standalone terminal coding agent, positioned directly against Anthropic's Claude Code and OpenAI's Codex CLI. It installs on macOS and Linux with a single shell command (Windows requires WSL2), and it ships alongside Muse Spark 1.2, a coding-focused update of Meta's foundation model that was co-trained with the agent harness.
The model brings a 1M-token context window, which is the practical reason Meta pitches Muse Code at large, legacy codebases: dependency graphs, thousands of files, and long debugging sessions in one session.
How it works
Muse Code runs specialized background agents that stay active for the whole session, building up context over time instead of starting fresh on each task. When a job is big enough, it fans out to separate sub-agents working in parallel, each in its own isolated git worktree so the parent working copy is never touched. Concurrency scales with CPU cores, clamped between 2 and 16 child agents.
Every action is logged as plain JSONL under ~/.local/share/muse/sessions/, which supports crash recovery and replay. Safety controls include approval modes for dangerous shell commands and an OS-enforced sandbox (Seatbelt on macOS, bubblewrap on Linux).
- Bundled skills: /plan (grounded planning), /grill (decision-forcing design interviews), /taste (an anti-slop filter for UI work)
- Project rules in AGENTS.md, with CLAUDE.md fallback for compatibility
- Skill migration from other agents: muse skills import --from claude or --from codex
- Reasoning effort dial from none to ultra via /effort
Muse Code and MCP
Muse Code supports the Model Context Protocol in the shipped beta. MCP servers are configured in ~/.config/muse/settings.json under an mcp_servers key, with two transports: stdio for local processes and streamable_http for remote servers. That is how Muse Code reaches tools beyond the repository, such as email, calendar, documents, and spreadsheets.
One line in Meta's own documentation deserves attention: MCP tools are not sandboxed. Unlike shell commands, an MCP server runs outside Muse Code's filesystem and network sandbox. Whatever access an MCP connection grants, the agent holds it without the sandbox in the way, so the scoping has to happen on the server side of the connection.
What to scope before connecting business data
A coding agent with email or drive access is convenient: it can read bug reports, pull release notes together, and check a spreadsheet of test accounts. It also concentrates risk, because Muse Code's sub-agent fan-out means every child agent inherits the same connections the session holds, and Meta's contributor pricing tier, which cuts token prices by roughly an order of magnitude, comes with training rights on your data (the standard tier does not; check Meta's current terms for your tier).
The practical fix is the same one that applies to any AI client: connect data through a layer that enforces least privilege per token, redacts sensitive fields before the model sees them, and records every tool call. That keeps a debugging session from becoming an open door to the whole inbox.
- Muse Code is Meta's terminal coding agent (beta, August 2026), powered by Muse Spark 1.2 with a 1M-token context window.
- It fans out to parallel sub-agents in isolated git worktrees; all of them inherit the session's data connections.
- MCP is supported via settings.json with stdio and streamable_http transports.
- Meta documents that MCP tools run outside Muse Code's sandbox, so access control must live server-side.
- Meta's discounted contributor API tier includes training rights on your data; the standard tier does not.
Frequently asked questions
What is Muse Code in one sentence?
Muse Code is Meta's terminal-based AI coding agent, released in beta in August 2026, that plans, writes, and validates code across large repositories using the Muse Spark 1.2 model.
Is Muse Code free?
Meta offers consumer subscriptions starting at $5 per month and pay-as-you-go API pricing at $1.25 per million input tokens and $4.25 per million output tokens on the standard tier. A discounted contributor tier exists at roughly a tenth of the price, but it grants Meta training rights on your data. Pricing changes fast in beta, so confirm current terms with Meta.
Does Muse Code support MCP?
Yes. The shipped beta supports MCP servers configured in ~/.config/muse/settings.json, with stdio and streamable_http transports. Some launch-week articles claimed MCP was missing; that was incorrect.
What platforms does Muse Code run on?
macOS and Linux on x86_64 and arm64, installed with a single curl command. There is no native Windows build; Windows users need WSL2.
How is Muse Code different from Claude Code?
They are the same category of tool: a terminal coding agent driven by a frontier model. Muse Code's launch distinctives are automatic sub-agent fan-out into isolated git worktrees, persistent background agents, a 1M-token context window, and Meta's aggressive pricing. Claude Code has the more mature skill and MCP ecosystem. Both can reach business data through MCP, and both deserve the same least-privilege scoping when they do.
Is it safe to connect email or drive to Muse Code?
It depends entirely on the scope of what you connect. Meta's docs are explicit that MCP tools run outside Muse Code's sandbox, and every parallel sub-agent inherits the session's connections. Connecting through a data firewall that enforces per-token permissions, redacts sensitive fields, and logs every tool call reduces the risk to the workflow you actually intended.
Keep exploring
PortEden is a software provider, not a law firm, accounting firm, or compliance auditor, and nothing on this page is legal, compliance, tax, or other professional advice. PortEden does not issue compliance certifications, attestations, or audit opinions. This content is provided for general informational purposes only, on an as-is basis and without warranties of any kind, and may not reflect the most current laws, regulations, or your specific situation. Before acting on it, consult a qualified attorney, auditor, or compliance professional.