Muse for developers: API, Muse Code, Glimmer and connectors

Developers can build with Muse at four layers: the Meta Model API for Muse Spark, the Muse Code terminal agent, the open-weights Muse Glimmer 30B model, and connectors that plug your service into the Muse personal agent. The API is drop-in compatible with the OpenAI and Anthropic SDKs, so most existing code works by changing a base URL and key. Glimmer is Apache 2.0 and runs on a 24–32GB machine.

How do I call Muse Spark from code?

The Meta Model API lives at https://api.meta.ai/v1 and accepts the Responses, Chat Completions and Messages formats. Auth is a Bearer MODEL_API_KEY.

What Spark 1.3 supports, per Meta's docs:

• 1,048,576-token context
• parallel tool calls with streamed arguments
• structured output and prompt caching
• text, image and video/document input

Available model IDs include muse-spark-1.3, muse-spark-1.3-contributor, muse-image-1.0, muse-voice-transcribe-1.0 and sam-3.1.

One wrinkle: Spark emits a non-standard SSE event that trips some OpenAI-compatible clients. Muse Spark Anywhere documents fixes for pi, opencode and others.

What does the Muse API cost?

From Meta's pricing page:

Tier	Input / M	Output / M	Cached / M	Trains on data?
Standard	$1.25	$4.25	$0.15	No
Contributor	$0.10	$0.20	$0.002	Meta may

Extras: web search grounding at $2.50 per 1,000 queries, Muse Image at $0.01 per image, voice transcription at $0.18 per audio hour. Rate limits per team are 3,000 RPM / 4M TPM on standard and 100 RPM / 3M TPM on contributor. Some third-party sites mention $20 in free credits for new accounts, which Meta's docs do not confirm.

You can also reach Spark through Vercel AI Gateway, OpenRouter and Cursor.

What is Muse Code?

Meta's terminal coding agent, co-trained with Spark, with subagents in git worktrees and replayable event logs (Meta). Plans run $5, $15 and $50 a month. The catalog's tools and apps section is full of Muse Code add-ons: editor bridges like muse-acp, guardrails like Maestro, and a Docker sandbox kit. The full rundown is in Muse Spark as a coding agent.

How do I run Muse Glimmer locally?

Glimmer 30B is a dense ~29.6B model with a 1.8B vision encoder, 131K context and Apache 2.0 weights (model card). Meta says it fits under 20GB at 4-bit. Day-one support landed in Ollama, LM Studio, llama.cpp, MLX, vLLM, SGLang and more.

Quick routes from the catalog:

• ollama run muse-glimmer (Ollama)
• LM Studio, which reports 83.3% on its BionicBench v0.1
• Glimmer as a backend for Codex and Claude Code
• Mozilla.ai's local agent opened a real PR, fixing a bug offline
• Unsloth's free fine-tuning notebooks, which run on 24GB

Before production, read the Dell GB10 negative result and the Glimmer vs Qwen benchmark, where Glimmer passed 24 of 30 against Qwen 3.8's 30 of 30. Local results vary widely.

How do I build a Muse connector?

Two paths, per Meta's help page:

• Custom connectors. Muse builds a connector from any API, stores credentials in a secure vault, and Meta does not review it. David Singleton showed Muse writing its own integration code in the user's VM.
• Directory connectors. Submit at muse.ai/platform for functional, security and legal review. No SDK, API spec or developer terms had been published as of September 18 (CellCog).

In practice, builders ship an OpenAPI spec, an llms.txt or SKILL.md file and a bearer token. Templates:

• Cloudflare Workers starter
• Deployable connector template with a registry that generates routes, OpenAPI and llms.txt
• 150 one-paste connector skills that declare their allowed hosts
• muse-fileapi, a local file connector with two-phase writes and an audit log

What does the Muse agent's runtime look like?

Useful if you are building for it. Each user gets a Muse Secure VM; a separate Sentinel authority gates connector actions and network egress; credentials are surrogated so the model never holds them (Meta). Community digging adds detail:

• Claude Code inside a Muse VM reports 2 vCPUs, ~8GB RAM, 100GB disk and no inbound ports.
• Exporting the agent filesystem found the home at /home/hatch, with soul.md, identity.md and memory files.
• Mapping Muse's on-disk architecture documents nightly "dreams" jobs and per-goal crons.

These are community findings, not documented APIs, so do not build products that depend on them.

What are the rate limits and quirks to plan for?

Contributor's 100 RPM limit is low for production traffic, so use it for batch jobs and prototypes. Spark's token savings claim is task-dependent: DataCamp measured a net 12% cost increase across three coding tasks (DataCamp). Effort level changes quality a lot; Morgan Linton's VulcanBench run found serious issues at lower effort.

Which layer should I start with?

Goal	Start with
Add an LLM to an app	Meta Model API, standard tier
Private or offline inference	Muse Glimmer
Cheaper coding help	Muse Code or Spark contributor in OpenCode
Reach Muse users	A custom connector now, directory listing later

More in our engineering guide and Muse for founders. Built something? Submit it.

Frequently asked questions

Is the Muse API compatible with the OpenAI SDK?

Yes. Meta describes the Model API as drop-in compatible with the OpenAI SDK, the Anthropic SDK and OpenAI-compatible CLIs. Point the base URL at https://api.meta.ai/v1 and use your Meta key.

Is Muse Glimmer open source?

The weights are released under Apache 2.0 on Hugging Face. Muse Spark is closed weights, though Meta has said a Spark open-weights release is coming.

Is there a Muse connector SDK?

Not as of September 18, 2026. Developers build connectors from a standard API plus an OpenAPI spec and docs, and submit them at muse.ai/platform for review.

What is the difference between Muse Spark standard and contributor?

Contributor is far cheaper, at $0.10 input and $0.20 output per million tokens, with lower rate limits, and Meta may train on the data. Standard prompts are not used for training.