Muse Code is Meta’s agentic coding harness for the terminal: it plans tasks, edits files, runs commands inside a sandbox and connects MCP servers to turn the Muse Spark models into a working AI harness for developers in the terminal. It is built by Meta. As of 20 September 2026.
What is Muse Code?
Muse Code comes from Meta and, per its own documentation, is built specifically for the Muse Spark model family. The verified official pages do not document a public announcement date for Muse Code; per the changelog, 0.1.0 was the first published version, and as of 20 September 2026 Muse Code is at version 1.2.1 according to the changelog. It is installed via curl -fsSL https://dev.meta.ai/install.sh | sh (macOS, Linux) or irm https://dev.meta.ai/install.ps1 | iex on Windows, and started by running the native muse binary inside a project directory. The software is proprietary, and no public source-code repository is documented. Beyond interactive terminal use, Meta provides a TypeScript SDK (npm install @muse-code/sdk), which exposes its own protocol for programmable integrations via the muse serve command.
Muse Code as an AI harness
According to the documentation, Muse Code plans, edits and runs commands to complete a task in a project, backed by approvals and an operating-system sandbox active from the very first run. On first launch in an unfamiliar directory, the application explicitly asks whether to trust the workspace. Project instructions live in AGENTS.md, and Muse Code also reads an existing CLAUDE.md; user settings live in ~/.config/muse/settings.json. For memory, the documentation distinguishes personal project memory, stored outside the repository, from committed project memory shared with the team under .agents/memory/ – with an explicit warning that shared memory can act as an attack surface for prompt injection in untrusted workspaces. External systems connect via MCP servers over stdio or streamable_http, though MCP tools bypass sandboxing according to the documentation. Skills live under .agents/skills/<skill-id>/SKILL.md and are invoked via slash commands; hooks wire shell commands to lifecycle events such as SessionStart or PreToolUse via <project-root>/.muse/hooks.json. For complex tasks, a lead session spawns subagents with optional git worktree isolation, up to eight concurrently by default, configurable via agents.execution_capacity. Four observer agents, enabled by default, run alongside and propose actions on memory recall, skill use, goal tracking and verification of claimed work, without executing them directly.
Models and model access
According to its documentation, Muse Code is built specifically for Meta’s Muse Spark model family; the default model is muse-spark-1.2, selectable via the --model startup flag or mid-session via the /models slash command. A documented switch to other vendors’ models within Muse Code is not available. A freely configurable custom API endpoint for model access is not publicly documented. Whether model access can nonetheless be routed through a central proxy therefore remains open; organisations that want to budget and log Muse Code consistently alongside other model providers can use innFactory’s AI Gateway for the rest of their OpenAI-compatible model stack.
Enterprise use: data protection and governance
For the Muse Spark models, the Meta Model API documents a standard and a significantly cheaper contributor tier; on the contributor tier, users grant Meta the right to use prompts and outputs to train future models in exchange for reduced cost. This tier choice, per the documentation, also applies to Muse Code’s usage-based access, which sits alongside a separate flat-rate monthly subscription; the documentation names no dedicated EU or GDPR addendum. On the security side, the documentation explicitly notes that MCP tools bypass sandboxing and that hooks run outside the sandbox with a cleared environment – both should feed into an internal risk assessment. Important in practice: source code and prompts go to Meta as the model provider, which should be mapped in advance in an organisation’s records of processing activities, together with a check of current terms directly with Meta.
Orchestrating Muse Code with herdr
According to the documentation of the terminal orchestrator herdr, it detects Muse Code sessions via a screen manifest, but without state or session integration. This allows a basic status view of Muse Code instances alongside other coding agents in one shared terminal interface.
Where it fits: developer harness and enterprise harness
Muse Code is a harness for developers working on complex, often long-running coding workstreams with multiple subagents. For business users outside development, CompanyGPT is the matching harness: tools via MCP servers and n8n workflows, context via companyRAG, roles from Entra ID or Keycloak, and audit logs for business departments. The AI Gateway forms the shared cost and governance layer for both worlds, wherever the respective harnesses allow a custom endpoint. For an overview of other coding agents and orchestrators, see our overview of all AI harnesses. We help organisations evaluate and adopt new coding harnesses such as Muse Code too, securely, with proper governance and centralised cost control.