Muse Spark as a coding agent: the model behind Muse Code, Cursor and opencode

Muse Spark works as a coding agent by powering the tool you code in: Meta's own Muse Code terminal agent, or third-party harnesses such as Cursor, opencode, pi and Cline through the Meta Model API. Meta pitches Spark 1.3 as tuned for long-horizon coding, and its benchmark numbers are strong. Independent tests are more mixed, and the most popular community pattern treats Spark as a cheap worker with a stronger model checking its output.

Muse Spark vs Muse Code: what's the difference?

Muse Spark is the model. Muse Code is Meta's terminal coding agent, co-trained with Spark and launched in beta on August 5, 2026 alongside Spark 1.2 (Meta). Muse Code is one way to run Spark; Cursor, opencode, pi and Cline are others. This post is about Spark as the engine across all of them. For Muse Code itself, with its $5, $15 and $50 plans, install commands and settings, see our Muse Code page.

One detail matters whichever route you pick. Muse Code defaults to the contributor model (DataCamp), and on the API the contributor tier costs $0.10 per million input tokens and $0.20 output, against $1.25 and $4.25 for standard, but Meta may train on contributor data (pricing). If your code is private, check which model you are on.

How good is Muse Spark 1.3 at coding?

Meta's numbers, via MarkTechPost: 75.4 on DeepSWE v1.1 (Opus 5 at 74.0), 88.8 on Terminal-Bench 2.1, and 59.4 on SWE-Atlas Codebase QnA. Meta also claims about 20% fewer tool calls and 25% fewer tokens than 1.2.

Outside tests, as reported by their authors:

• DataCamp ran three real tasks. Two used 23–32% fewer completion tokens, but a refactor used 70% more, for a net 12% cost increase.
• Paweł Huryn's planted-bug test put Spark 1.3 (max) at 33 of 105 bugs fixed, tied with Fable 5.1 (high) and ahead of Opus 5 and Grok 4.6 at 27.
• Morgan Linton's VulcanBench run found serious issues at lower effort levels.
• eesel reports #6 on the Artificial Analysis Intelligence Index, trailing Opus 5 on four of six agent evals.
• MindStudio got "a cube shooting at other cubes" from an open-ended game-clone prompt.

The summary: effort level matters a lot, and the token savings are task-dependent.

Can I use Muse Spark inside my editor?

Several community bridges exist. None are official.

• muse-acp is a dependency-free Rust bridge to Zed and JetBrains IDEs using your Muse Code subscription.
• muse-code-acp is a Node adapter for Zed or VS Code with streaming and MCP servers.
• Muse Spark Code for VS Code runs Spark as an in-editor agent with reviewable diffs and permission modes.
• Helicon wraps the CLI in a desktop app with session history and diffs.
• Muse Code Remote lets you drive the terminal agent from any browser over an outbound WebSocket relay.

Which other harnesses run Muse Spark?

Spark is on Vercel AI Gateway as meta/muse-spark-1.3, on OpenRouter, in Cursor, and as a free contributor model in OpenCode. Builders filled the gaps:

• pi-meta-ai adds Spark to the pi agent.
• Muse Spark Anywhere patches Spark's non-standard SSE event so it works in pi, opencode and other OpenAI-compatible tools.
• Muse Code via OpenRouter runs the Muse Code harness on an OpenRouter key.
• Cline ported instructions from the Muse harness prompt when Muse Code would not sign in from Docker.

What is the cheap-worker pattern?

The most common serious setup in the catalog pairs a strong planner with Spark doing the bulk of the edits:

• sol-loop: a GPT planner, Spark executes each atomic spec and returns evidence. The author says over 99% of tokens run on Muse.
• ocodex: Spark 1.2 contributor workers fan out while one paid supervisor audits every claim.
• opencode-delegate and musecode: Claude Code plans and Spark 1.3 in opencode implements.
• delegate-to-muse: a skill that hands work to muse exec, then re-runs tests rather than trusting the summary.
• shunt-muse: routes large file reads to Spark. The author measured about 4,800 tokens out and 300 back on a 617-line file.

The shared idea is simple. Use the cheap model for volume, and verify with tests or a second model.

How do I keep a Spark coding agent from breaking things?

Guardrail plugins are one of the busier corners of the catalog:

• Maestro adds a command guardrail, proof-of-done checks and secret redaction.
• The Agentic Control Plane plugin policy-checks every tool call and logs allow, ask or deny.
• The Docker sandbox kit runs Muse Code in isolation with network access limited to Meta's services.

Is Muse Spark good for large refactors?

When it launched Spark 1.2, Meta pitched the 1M-token context for multi-file refactors and hours-long tasks (Meta). Independent data is mixed: DataCamp's refactor task used 70% more completion tokens than on 1.2. Start with a refactor that has good test coverage.

Where do I start?

Watch the freeCodeCamp course, read the Muse Code handbook field notes, then pick one small repo task. For more, see Muse for developers, our engineering guide and the tools and apps category.

Frequently asked questions

Can I use Muse Spark for coding for free?

Yes, with a catch. OpenCode offers Muse Spark 1.3 as a free contributor model, and the API's contributor tier costs $0.10 per million input tokens. Meta may train on contributor data, so keep private code on the standard tier. Muse Code, Meta's own agent, starts at $5 a month.

Does Meta train on code I send to Muse Spark?

Only on the contributor tier. Standard-tier prompts on the Meta Model API are not used for training. Muse Code uses the contributor model by default, so switch it to Standard if your code is private.

Is Muse Spark better than Claude for coding?

Meta's DeepSWE number edges Opus 5, but independent results are mixed: eesel reports it trailing Opus 5 on four of six agent evals, while one planted-bug test had Spark ahead. Test it on your own tasks at the effort level you plan to use.

Can I use Muse Spark in Cursor?

Yes. Cursor announced Muse Spark 1.3 support, and it is also available through Vercel AI Gateway and OpenRouter.