The Muse Developer Platform: APIs, Muse Code, Glimmer and Connectors

The Muse developer platform is four separate things: the Meta Model API for calling Muse Spark, Muse Code as a terminal coding agent, Muse Glimmer as open weights you host yourself, and the connector platform at muse.ai/platform for reaching Muse agent users. The first three are documented and priced. The connector platform has a review process and a directory, but as of its September 18 launch there was no SDK, API spec, developer terms or published revenue share.

Here is each piece, what it costs, and what builders have done with it.

The Meta Model API

The API lives at https://api.meta.ai/v1 and is "drop-in compatible" with the OpenAI SDK, the Anthropic SDK and OpenAI-compatible agent CLIs. It accepts Responses, Chat Completions and Messages formats, with a Bearer MODEL_API_KEY (docs). There is no dedicated Meta SDK to learn; you point an existing client at a new base URL.

Models: muse-spark-1.3, 1.2 and 1.1, the contributor variants muse-spark-1.3-contributor and muse-spark-1.2-contributor, muse-image-1.0, muse-voice-transcribe-1.0 and sam-3.1.

Pricing (docs):

Item	Price
Standard input / output / cached input	$1.25 / $4.25 / $0.15 per M tokens
Contributor input / output / cached input	$0.10 / $0.20 / $0.002 per M tokens
Web search grounding	$2.50 per 1K queries
Muse Image	$0.01 per image
Voice Transcribe	$0.18 per audio hour

Standard prompts are not used for training; contributor data may be. Rate limits are 3,000 RPM and 4M TPM per team on standard, 100 RPM and 3M TPM on contributor. Some third-party sites mention $20 in free credits for new accounts, which Meta's docs do not confirm.

Capabilities: a 1,048,576-token context, parallel tool calls with streamed arguments, structured output, prompt caching, tool search with deferred loading, adjustable reasoning effort, and text, image and video/document input. Details in structured outputs and tool calling.

Where else it runs: Vercel AI Gateway as meta/muse-spark-1.3 (Vercel), Cursor, OpenRouter, and OpenCode, which offers it as a free contributor model.

What people build on the API

• Tarquin, a voice-driven StackChan robot butler with Spark 1.3 as its brain.
• reZme, a SwiftUI Mac app that turns YouTube videos into reports.
• windows-cua, a native Win32 computer-use agent on the Responses API.
• MdChat, a Markdown editor with a Spark side panel.

Integration friction exists. muse-spark-anywhere patches a non-standard SSE event and missing catalog entries so Spark works in pi, opencode and Agent Orchestrator.

Muse Code, the terminal agent

Muse Code launched in beta on August 5 alongside Spark 1.2. It is co-trained with Spark, runs subagents in isolated git worktrees, keeps auditable and replayable event logs, and ships skills like /plan, /grilling and /taste (Meta).

• Install: curl -fsSL https://dev.meta.ai/install.sh | sh on macOS/Linux, irm https://dev.meta.ai/install.ps1 | iex on Windows.
• Plans: $5/month Everyday, $15 High (5x usage), $50 Power (20x usage) (dev.meta.ai).
• Default model: the contributor model, per DataCamp.

The ecosystem around it is mostly bridges. Muse Code speaks its own Muse Session Protocol, and builders have wrapped it for Zed and JetBrains, VS Code, a desktop app, a browser remote and a Docker sandbox. Plugins add guardrails: Maestro redacts secrets and tracks cost, and the Agentic Control Plane plugin policy-checks every tool call. The freeCodeCamp course is the best long-form tutorial.

Muse Glimmer, open weights

Glimmer 30B is Apache 2.0, dense, about 29.6B parameters with a vision encoder, 131K context, and 100+ languages (model card). The Model API does not serve it; you run it yourself or through Together, Fireworks, OpenRouter or NVIDIA's hosted endpoint.

Day-0 runtimes included Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM and SGLang, with AMD, Arm, Dell, Intel and NVIDIA as hardware partners. Hardware reports are collected in running Glimmer on a consumer GPU, and training in fine-tuning Glimmer. Meta says a "Muse Spark open weights release" is also coming, without a date (Meta).

The connector platform and directory

This is the piece that reaches consumers. On September 18, Zuckerberg opened Muse connectors to developers at muse.ai/platform. The flow has three steps:

1. Describe your connector.
2. Review: Meta checks it against "functional, security, and legal" requirements and runs end-to-end testing.
3. List: approved connectors appear in the Muse directory, with possible editorial featuring.

Stripe Link handles payments. Meta received more than 1,500 applications in under a week (TechCrunch).

What was missing at launch, per CellCog: an SDK, an API spec, developer terms, a review timeline, fees and revenue share. Plan accordingly if you are betting a product on the directory.

Directory company is heavy. Launch connectors included Gmail, Google Calendar, Plaid, Stripe Link, Shopify Catalog, OpenTable, Ticketmaster and Duffel, and Connect 2026 added Walmart, Best Buy, Sephora, Notion, Granola, GitHub, Box and others (Meta).

Distribution without the directory

You do not need Meta's approval to reach Muse users. Any Muse user can ask the agent to build a custom connector from your API, and Meta does not review those (Meta help). Credentials stay in Muse's secure store behind surrogate tokens.

Builders have made this route easier with starters:

• Cloudflare Workers starter: one bearer token in front of upstream keys, an OpenAPI spec, a SKILL.md template and a proof checklist.
• Deployable connector template: per-connection tokens and async jobs, generating routes, OpenAPI and llms.txt for Vercel, Workers or Docker.
• Threadwell Slack: an example submission pack for the directory.

Schedulers like Postiz already distribute this way with a paste-in setup prompt. For how MCP fits, see MCP servers vs Muse connectors.

Security considerations for developers

Muse runs each user in a dedicated VM, gates connector actions and network egress through a separate Sentinel authority, and pays bug bounties up to $300,000 (Meta research). Your connector still has to earn trust on its own: declare allowed hosts, scope tokens, default to read-only. The 150 one-paste skills project declares allowed hosts per skill, which is a good habit to copy.

Browse tools and apps for more developer builds, or submit yours.

Frequently asked questions

Is there a Muse SDK?

Not a dedicated one. The Meta Model API works with the existing OpenAI and Anthropic SDKs by changing the base URL. For connectors, no SDK or API spec had been published as of the platform's September 18 launch.

How much does the Meta Model API cost?

Standard Muse Spark pricing is $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15. The contributor tier costs $0.10 and $0.20 but Meta may train on that data.

How do I publish a connector to the Muse directory?

Submit it at muse.ai/platform. Meta reviews it for functional, security and legal requirements and runs end-to-end tests before listing it. No review timeline, fees or revenue share have been published.

Can I build a Muse connector without Meta's approval?

Yes. Any Muse user can have the agent build a custom connector from your public API or MCP server, and Meta does not review custom connectors. Publishing a clean OpenAPI spec and a setup prompt is the fastest way to reach users.

Does the Meta Model API serve Muse Glimmer?

No. Glimmer is self-hosted. You can run it locally with Ollama, LM Studio or llama.cpp, or use third-party hosts such as Together, Fireworks, OpenRouter or NVIDIA's build.nvidia.com endpoint.