Meta MusePersonal agents, models, and developer tools.
Muse handles tasks across connected services. Muse Code works in your repository. Muse Spark powers the reasoning. Explore how the products fit together, what they cost, and where permissions apply.
Personal Muse: rolling out in the US.
Which Muse product does what?
Start with the task. Choose a ready-made agent, call a model through the API, or run a model on your own hardware.
Use the app or web interface to delegate a task. A cloud computer runs the tools and retains the working context.
From a request to an action
Personal Muse can browse, work with email and calendars, and continue a task after you close the app. Give it an outcome, constraints and an approval boundary.
Find a New York–Tokyo round trip next month under $1,200. Watch prices for seven days and ask me before buying.
- 01
Set constraints
Dates, route and budget
- 02
Find options
Search and compare flights
- 03
Monitor prices
Recheck over seven days
Request approval
Ask before a purchase
Where permissions are enforced
A dedicated Linux VM runs the agent. Sentinel controls connector actions and outbound requests; authd stores credentials outside the agent's runtime.
- Requests can be allowed, denied, or sent to you for approval.
- The agent uses surrogate tokens; real credentials are inserted after authorization.
Isolation has limits
Secure VM does not prevent Meta from accessing data to operate the service. Confidential VM is planned. Prompt injection remains an open problem.
Check data-use settings
Personal Muse can use sanitized interactions for training; settings provide an opt-out. API Standard and Contributor have separate data terms.
Muse Code and Meta Model API
Muse Code is the ready-made agent for a terminal or CI pipeline. Meta Model API gives your own application access to the models.
Muse CodePlans, edits files and runs commands in your project.
Meta Model APIhttps://api.meta.ai/v1
curl -fsSL https://dev.meta.ai/install.sh | sh
cd /path/to/your/project
museFirst run asks you to sign in and trust the project.
Conceptual multi-agent workflow. Availability of staged workflows varies by build. Windows currently lacks session messaging and voice input.
The models behind the products
Spark returns text and tool calls. Image generates pictures. Voice Transcribe turns audio into text. Glimmer is downloaded and self-hosted.
Muse Spark 1.3
Coding, tool use and long-running agent tasks.
1,048,576tokens of contextmuse-spark-1.3Text, image, video and PDF inputs. Audio quality is limited in 1.3; use Spark 1.2 or Voice Transcribe for audio.
Muse Image
Generate and edit images.
$0.01 / imageMuse Voice Transcribe
Speech recognition with speaker diarization.
$0.18 / hourMuse Glimmer
30B open-weight model for local agents. Hardware requirements depend on the runtime and quantization.
Apache 2.0SAM 3.1
Locate and segment objects in images and video.
Boxes & masksServices Muse can connect to
A selection from Meta's launch and Connect announcements. Announced partnerships may roll out gradually; check your account for availability.
Bring your service to Muse
Describe your connector → submit it for functional, security and legal review → complete testing → appear in the directory if approved.
Subscriptions and API usage are billed separately
These subscriptions are for Muse Code. API estimates below cover Spark token charges only, in US dollars.
Everyday Usage
$5/ month
About 10–50 prompts / 5 hours
High Usage
$15/ month
5× the Everyday usage allowance
Power Usage
$50/ month
20× the Everyday usage allowance
Muse Spark 1.3 API cost calculator
| Tier | Uncached input | Cached input | Output |
|---|---|---|---|
| Standard | $1.25 | $0.15 | $4.25 |
| Contributor | $0.1 | $0.002 | $0.2 |
Enter total tokens for your estimate. Count cached and uncached input separately, without overlap.
Token charges only; excludes other tools, taxes and subscriptions.
Prompts and completions are not used for training.
Allows training on prompts and completions; max reasoning is unavailable.
Hardware announcements
Muse on AI glasses
Meta says the integration will arrive in the coming months, allowing the agent to respond to what you are looking at.
Muse Charm
A pocket-sized voice device was announced at Connect. Meta says more details will follow later this year.
Common questions
Is Muse Code the same as Muse Spark?
Spark is the model. Muse Code adds the project environment, tools, command execution and approvals needed for coding work. You can also call Spark from your own software.
Is Contributor simply a cheaper Standard plan?
No. Contributor permits training on prompts and completions, and the max reasoning level is Standard-only. Choose according to data requirements as well as price.
Can I run Muse locally?
Personal Muse uses a cloud VM. Glimmer is a separate open-weight model for your hardware; it does not provide the complete personal Muse service by itself.
Is this a hands-on review?
This guide summarizes product documentation. Diagrams simplify the documented architecture; the sample task flow is illustrative. No performance or success-rate claims are based on our own testing.