Muse Code
Meta's coding agent for the terminal and CI, run as muse and built for Meta's Muse Spark models, which Meta says it co-trained with the agent. Meta released it in beta on August 5, 2026 alongside Muse Spark 1.2. It plans, edits, and runs commands with approvals and an OS sandbox on from the first run, coordinates multiple agents by default (parallel workers with reviewers in the background), can operate browsers, desktop, and mobile apps through computer use, and supports subagents, skills, hooks, and MCP. Parallel and staged Workflows are still rolling out and are not on every build or platform. Programs can drive a session over a versioned protocol (muse serve) with a TypeScript SDK; the SDK is MIT licensed on GitHub, while the agent itself ships as a native binary. One caution from Meta's own docs: it reads a repository's committed MEMORY.md into the model's context even in an untrusted workspace, so treat that file as a prompt-injection surface on code you do not control. On the Terminal-Bench 4.0 board read on October 28, 2026, Muse Code with Muse Spark 1.3 at xhigh effort scored 14.5 percent.
Leaderboard Placements
Live board as of 2026-10-28: federated Terminal-Bench rows, with TensorFeed editorial scores for harnesses the board does not cover.
| Benchmark | Best base model | Score | Rank |
|---|---|---|---|
| SWE-bench Verified | - | - | - |
| Terminal-Bench | - | - | - |
| Terminal-Bench 2.1 | - | - | - |
| Terminal-Bench 3.0 | - | - | - |
| Terminal-Bench 4.0 | Muse Spark 1.3 | 14.5 | #19 / 21 |
| Aider Polyglot | - | - | - |
| SWE-Lancer | - | - | - |
Distribution
Native binary for macOS, Linux, and Windows, installed with curl -fsSL https://dev.meta.ai/install.sh | sh or the PowerShell installer. Voice input and session messaging are unavailable on Windows. CI runs use a META_API_KEY with the headless muse exec mode.
Model Story
Built for Muse models, with muse-spark-1.2 as the default. Override it per run with the model flag or mid-session with the /models command. The plans cover Muse Spark, Muse Wonder, and Muse Image, and the Power plan adds Muse Wonder Full.
Pricing
Monthly plans at $5 (Everyday Usage), $15 (High Usage), $50 (Power Usage), and $200 (Maximum Usage), each scaling the prompt allowance. Or pay per token through the Meta Model API: Muse Spark 1.3 is $1.25 input, $0.15 cached, and $4.25 output per million tokens, or $0.10 and $0.20 on the Contributor tier, whose data Meta uses to improve its products.
Who It's For
Developers on Meta's Muse models who want a multi-agent terminal agent with sandboxing on by default, and teams wiring an agent into CI or their own tooling through the session protocol.
Notable Features
- Multiple agents by default: parallel workers with background reviewers
- OS sandbox and approval gating on from the first run
- Computer use across browsers, desktop, and mobile apps
- Headless muse exec for scripts and CI
- Versioned session protocol with an MIT-licensed SDK