Muse Code + Muse Spark 1.2: Meta's Terminal Coding Agent
On 5 August 2026 Meta Superintelligence Labs shipped Muse Code — a terminal coding agent — alongside Muse Spark 1.2, the model co-trained with it. Two products, one release, and the pairing is the interesting part: Meta claims tools that are trained together with the model produce fewer retries and higher-quality output than a generic model steered by a generic harness. Whether or not you buy that claim, the launch is worth reading closely — Muse Code is the closest structural analogue to Claude Code that any competitor has shipped, and a few of its design decisions (the append-only event log, the contributor tier pricing, the three bundled skills) are non-obvious enough to be worth stealing regardless of which agent you actually use day to day.
This page is the practical read: what Muse Code actually is, how it compares to Claude Code on the numbers Meta itself published, the pricing trap that shows up on the second read of the pricing table, and where it fits into the AILmanac tour.
- Understand what Muse Code is architecturally: a terminal agent with a local append-only event log, persistent async background agents, and three shipped skills (/plan, /grill, /goal)
- Read the Terminal-Bench 2.1 and DeepSWE 1.1 numbers correctly — including the vendor-run caveat that the source review itself flags
- Know the pricing tiers: $1.25/$4.25 standard vs. $0.10/$0.20 contributor, and what the 12.5×/21.25× discount actually costs you in data terms
- Get the install command, the platforms supported, and the beta-status caveats before you point this at anything sensitive
- Place Muse Code next to Claude Code, Codex CLI, and the rest of the coding-agent CLI landscape
The one-sentence version
Muse Code is a beta terminal coding agent for macOS and Linux, powered by Muse Spark 1.2 — a 1M-token multimodal-input, text-output model — with async background agents, a local append-only event log for crash-safe resume, and three bundled skills (/plan, /grill, /goal); Meta prices the standard tier at $1.25/M input and $4.25/M output, and offers a contributor tier at $0.10/M input and $0.20/M output in exchange for permission to train future Meta models on your prompts and completions.
Four things that aren't obvious from the launch headlines
1. The event log is a write-ahead log, and it's what makes long runs safe
Muse Code's runtime records every model call, tool run, approval, and edit to a local append-only event log before it executes. Meta describes this as "replay-exact and restart-safe." If the agent crashes mid-task — segfault, laptop lid, expired token, anything — you resume from the log without losing work or having to re-prompt the model to reconstruct state.
That's a write-ahead log pattern from database engineering, applied to the harness rather than the storage layer. Two things follow that are worth internalising:
- Failures during 24-hour runs become recoverable rather than expensive. Meta's own kernel-optimisation case study executed 1,000+ tool calls over 24 hours on NVIDIA Hopper GPUs. Without a durable log, one crash in hour 22 costs the whole run. With one, the model resumes from the last committed event.
- Post-hoc auditing is trivial. Every decision is on disk, in order, with the tool inputs and outputs. You can inspect what the agent thought, why it called a particular tool, and what came back — without instrumenting anything.
The comparable pattern in Claude Code is checkpoints and rewind, which snapshots the workspace so you can step backwards through interactive edits. The Muse Code event log is a different tool for a different problem: rewind lets you undo, while the event log lets the agent resume.
2. The contributor tier is a 12.5× / 21.25× discount, and it's not free
Two pricing tiers, publicly disclosed:
| Tier | Input ($/M) | Cached input ($/M) | Output ($/M) | You give up |
|---|---|---|---|---|
| Standard | $1.25 | $0.15 | $4.25 | Nothing |
| Contributor | $0.10 | — | $0.20 | Permission to train future Meta models on your prompts and completions |
The math on the discount: 12.5× cheaper on input, 21.25× cheaper on output. For a real coding session that reads more than it writes, the effective cost drop lands closer to 15×.
The catch is not that Meta will look at your prompts (all frontier vendors log for abuse review by default). The catch is training rights — Meta explicitly retains permission to train future models on your inputs and completions. That means:
- Anything that would leak trade secrets, PII, licensed code, or a customer's data through a future model output is off-limits.
- Anything under a share-alike or similar restrictive licence you don't own is off-limits.
- Anything covered by an NDA is off-limits.
Read Meta's exact terms before flipping the toggle — the source reviews are clear that "confirm exact terms and toggle settings in Meta's official docs before running it on anything sensitive." A 12.5× cost reduction on personal side-projects is a genuinely good deal. A 12.5× cost reduction on your employer's monorepo is a legal problem waiting to happen.
3. Meta's own charts place Muse Spark 1.2 second, behind Claude Opus 5
Meta published Terminal-Bench 2.1 and DeepSWE 1.1 comparisons that put Muse Spark 1.2 behind Claude Opus 5 on both. That's an unusually honest launch chart, and it's worth reading precisely rather than dismissively.
| Benchmark | Muse Spark 1.2 | Claude Opus 5 (paired with Claude Code) |
|---|---|---|
| Terminal-Bench 2.1 | 82.9% | Higher — Meta ranks it first (external secondhand reporting cites ~86.7%) |
| DeepSWE 1.1 | 59.3% | Higher — Meta ranks it first (score not published in the reviews we could verify) |
Two caveats before you take those numbers as gospel:
- Vendor-run methodology. The published numbers are pass@1 averaged over five attempts, each model paired with its own agent product, and Meta itself notes its setup "may not be tuned for third-party models." All numbers are vendor-run, not independent reproductions.
- Model+harness is the unit of measurement. Terminal-Bench doesn't score a bare model — it scores a model working through an agent. Muse Spark 1.2 is being evaluated inside Muse Code; Opus 5 inside Claude Code. The result is a claim about the pairing, not the model in isolation. Swapping either half changes the number.
The practical read: Muse Spark 1.2 is a strong second on the two benchmarks Meta chose to publish, in the harness Meta co-trained it for. That's a defensible position for a v1.2 model against a mature Claude Code / Opus 5 pairing — but "second on the vendor's own chart" is also exactly as strong a claim as it sounds. Don't over-extrapolate.
4. The three bundled skills are worth stealing even if you never install Muse Code
Muse Code ships three skills out of the box, and the shape of each is instructive:
- Converts a task into a concrete plan and pauses for human approval before execution. Same pattern as Claude Code's plan mode (docs/claude-code/plan-mode), and the reason is the same: the model is much better at proposing structure than at executing structure it never showed you first.
- Runs the plan through adversarial questioning until it stabilises — 'what breaks if X?', 'what did you assume about Y?', 'is there a case where Z fails silently?'. This is the missing step in most naive agent loops: not planning, but pressure-testing the plan before you spend tokens executing it. Nothing about /grill is Meta-specific; you can implement the same discipline with a subagent in any coding CLI.
- Works toward a specified objective across many tool calls without asking for permission at every step. Pairs with the event log: the human authorises the goal, the log records the trajectory, and the agent runs. The 1,000+ tool call kernel-optimisation case study is a /goal run.
The interesting design pattern here is the plan → grill → goal pipeline: propose, adversarially critique, then execute under a durable log. That decomposition maps cleanly onto workflows you might already be running with Claude Code plan-mode plus subagents — and it's a cleaner mental model than the ad-hoc "keep going until it works" loop most people end up with.
Install and first run
Availability: macOS and Linux, public beta as of 5 August 2026. No Windows build at launch. Weights are hosted-only — Meta has not published downloadable weights for Muse Spark 1.2.
Install Muse Code (macOS / Linux beta)
curl -fsSL https://dev.meta.ai/install.sh | bash
Standard security hygiene applies to any curl | bash installer: at minimum, review the script before piping it to a shell. Meta hosts the installer under its own domain, but nothing about the install pattern makes the script inherently safer than any other.
- Beta software. Expect breaking changes, undocumented edges, and quota surprises before general availability.
- Contributor tier trains on your data. Do not flip it on for work code, client code, licensed code, or anything under NDA.
- Vendor-run benchmarks. Muse Spark 1.2 is second to Claude Opus 5 on Meta's own charts — the numbers to trust are your own on your own repo.
How it compares to Claude Code, at a glance
| Axis | Muse Code + Muse Spark 1.2 | Claude Code + Claude Opus 5 |
|---|---|---|
| Terminal harness | Yes (macOS, Linux beta) | Yes (macOS, Linux, Windows) |
| Context window | 1M tokens (input), text-only output | 1M tokens, up to 128K output |
| Adversarial planning skill | /grill (bundled) | Compose with subagents + plan mode |
| Approval-gated planning | /plan (bundled) | Plan mode |
| Crash-safe long runs | Append-only event log (replay-exact, restart-safe) | Checkpoints and rewind (workspace snapshot, human-driven undo) |
| Sandboxing | Parallel work isolated in separate Git worktrees | Worktrees plus permissions |
| MCP support | Not documented in the launch materials we could verify | Full MCP support |
| Cheapest legitimate tier | $0.10/M input, $0.20/M output (contributor tier — data training) | Prompt caching discount without giving up training rights |
| Benchmark ranking (vendor-run) | 2nd on Terminal-Bench 2.1 (82.9%), 2nd on DeepSWE 1.1 (59.3%) | 1st on both (per Meta's chart) |
For the broader field — Codex CLI, Aider, and the rest — see coding agent CLIs compared.
When to actually reach for it
Reach for Muse Code when:
- Your workflow is dominated by long, unattended runs where the event log's crash safety is more valuable than any specific tool integration. The 24-hour kernel-optimisation case study is the shape.
- You want a second, structurally different opinion on a task. A different model + a different harness is a genuinely different second opinion. The best use of a competitor CLI is often to run the same task through it and compare.
- You are a solo developer working on side-projects where the contributor tier's data-training trade-off is acceptable, and 12.5×/21.25× cheaper tokens materially change what's affordable.
Don't reach for it when:
- You need MCP tool integration, which is not documented in the launch materials.
- You need Windows support, which does not exist at launch.
- You are working on anything covered by an NDA, a restrictive licence, or contains client / employer / PII data, unless you can pay the standard tier and are certain no contributor-tier toggle is on.
- Your quality bar is "beat Claude Opus 5 on Terminal-Bench 2.1", because on Meta's own numbers it doesn't.
Check yourself
0/4Where it sits next to what you know
If you've read coding agent CLIs compared, Muse Code drops in as a new entry with the same shape as Claude Code and Codex CLI — a terminal harness that owns file edits, tool calls, and long-running agent loops. What's different is the runtime discipline (event log) and the tier structure (contributor pricing). If you've read the Opus 5 field guide, the direct point of comparison is that Muse Spark 1.2 is a credible-but-second competitor to Opus 5 on Meta's own numbers, in the harness Meta trained for it.
For the broader picture — how to choose across the field — start with choosing a model and what AI costs across providers. If you want to compare techniques rather than products, the plan → grill → goal decomposition is worth composing yourself in Claude Code with subagents and plan mode — you get the discipline without the beta and without the contributor-tier question.
Sources & further reading
- Introducing Muse Code and Muse Spark 1.2 — Meta AI Research — the official launch post: architecture, benchmarks, and the pairing thesis.
- Meta Ships Muse Code Coding Agent With Co-Trained Muse Spark 1.2 Model — Unite.AI — installation, pricing tiers, and context on the co-training claim.
- Meta AI Releases Muse Code (Beta) — MarkTechPost — event log design, async background agents, and the 1,000+ tool-call kernel-optimisation case study.
- Muse Code Beta — explainx.ai — pricing walkthrough, contributor-tier data-policy caveats, and worktree-based sandboxing.
- Introducing Muse Code and Muse Spark 1.2 — Simon Willison — the practical two-tier pricing highlight and the pelican-on-bicycle progression from Spark 1.1 to 1.2.
- Muse Spark 1.2 — Benchmarks, Specs & Release Date — Vorp Labs — the confirmed pricing table, context window (1,048,576 tokens), and the vendor-run pass@1 caveat on the benchmark numbers.
- Meta debuts Muse Spark 1.2 and first coding agent — Yahoo Finance — the competitive framing against OpenAI and Anthropic.