Skip to content
Muse Atlas

How Muse works

Start with the simple view, then follow the model, tools and controls underneath it.

The simple view

1

Ask

Start with a goal.

Describe what you want to accomplish and provide the information Muse needs. Set boundaries before work begins.

Help me improve this website and check that it works.
Your request

Improve this website.

Project filesDesign notes
Boundary: prepare changes for review.
2

Understand

Make sense of the task.

The model considers the goal, relevant context and constraints. A larger task can become several smaller pieces of work.

Read the files, understand the layout, and identify what needs to change.
GoalA clearer website
ContextFiles + design notes
PlanInspect → change → check
3

Work

Use permitted tools.

The surrounding software gives the model access to tools. Each action is limited by available capabilities and configured permissions.

Change the code, then open the result in a browser.
Concept: code, files and terminal in a working environment
BrowserFilesCode
4

Check

Review the outcome.

Inspect the result, run relevant checks, and correct problems. A person still decides whether the output meets the goal.

Review the page, the changes and the checks before accepting the result.
Review the work
✓ Inspect the rendered page
✓ Run relevant checks
↻ Correct remaining problems
Ready for your reviewIllustrative checks, not a live run.

What actually happens

A model does not operate software on its own. The application supplies context, executes tools and controls access to the work environment.

Your request passes to Muse / Spark for reasoning, context and planning, then browser, file and code tools, and result and review. Connected services require permissions. Review feeds back into planning.
Conceptual flow. The model proposes work; software executes tools and enforces configured controls. Steps may repeat.

Separate responsibilities, working together.

The model reasons

Interprets information and proposes the next step.

The software executes

Runs tools, manages state and returns observations.

The controls set limits

Restrict access and request approval where configured.

Advanced architecture

Context and memory

The application decides what task history and saved context are available across steps and sessions.

Other agents

Some workflows delegate bounded tasks to separate workers, then review and combine their results.

Credentials

A credential service can authorize access without exposing the underlying secret to the model.

Approvals

Sensitive actions can pause at a permission check before software executes them.

Network restrictions

Runtime and workspace configuration determine which external systems tools may contact.

Untrusted content

Webpages and files can contain hostile instructions. Controls need to distinguish source content from the user’s authority.

See the system in a development workflow.

Muse Code puts the model alongside a repository and development tools, with checks and review around changes.

Explore Muse Code