MCP lets a compatible AI agent inspect and change a Misaki Studio project through structured tools. It is a second route into the product—not a separate project format and not a screen-recording macro that imitates every click.
The visual editor and the MCP server use the same underlying project and file models. A Component created by an agent appears in the File Explorer. An element changed in the editor can be read by an agent. Generated code comes from the same Component data whichever route authored it.
The production MCP surface covers the main project areas, including:
projects, folders, and files;
Page, Component, Mockup, and Code Component authoring;
element trees, text, paths, images, Component instances, and other supported nodes;
styles, responsive screens, variants, properties, states, events, and logic;
timelines, keyframes, and supported media layers;
design-system tokens and themes;
APIs, storage, Content, databases, Documents, Blogs, Scripts, and environments;
project libraries, diagnostics, previews, screenshots, measurement, and code export;
local/Git export configuration and other shipping operations where their prerequisites are available.
There is deliberately no fixed tool total in this guide. The registered surface changes as the product changes, and the connected client's current tool schemas are the authority for exact names and parameters.
Coverage: UI + MCP, with workflow-specific exceptions.
The editor may use a canvas drag, inspector field, context menu, or file tab. MCP may reach the same stored result by passing a file UUID, node ID, and structured value to a tool. These routes are equivalent when their result, validation, persistence, and recovery behavior match—not merely because the labels sound similar.
Some editor state is intentionally not project content:
which file tabs are open and their order;
Previous/Next file-navigation history;
expanded Explorer rows;
panel width, zoom, and the currently visible inspector.
An agent can change the intended file without reproducing those navigation details. Conversely, MCP provides machine-readable reads such as preview DOM, computed styles, measurements, interpreter state, and diagnostics that do not need a matching editor button.
Each MCP session starts without a selected project. The agent must list the available projects and select one before it uses project-scoped tools.
That selection is independent of the project visible in the desktop window. This is useful when an agent works in one project while you review another, but it also means “change this project” is too vague. Name the project in the prompt and ask the agent to confirm the selection.
Read Select a project safely before the first write.
Use the agent as an inspectable collaborator:
Name the project and the intended file or element.
Ask it to read the current structure before changing anything.
Describe the result, constraints, and what must remain unchanged.
Let it perform the smallest coherent mutation.
Require a read-back, preview, screenshot, diagnostic, generated-code check, or interaction test appropriate to the change.
Review the result in the editor.
Refine it or undo the file-content change if necessary.
For a new visual build, the server itself instructs agents to read its current build playbook before editing, construct planned subtrees efficiently, and verify both behavior and generated output. You do not need to memorize that internal playbook; you should still ask for evidence of the result.
MCP and the UI share undo/redo for supported design-file content transactions. That does not make every tool call undoable.
undo and redo act on a referenced design file's content history.
Creating, moving, renaming, duplicating, or deleting whole files uses project/file operations rather than that Component content stack.
delete-file explicitly cannot be undone.
Export, Git, package publication, website publication, and remote synchronization affect destinations outside the selected file's content history.
Ask for confirmation before destructive or external actions, and read the workflow-specific guide instead of assuming Command/Ctrl+Z can recover them.
The desktop editor can display MCP activity on affected file tabs, File Explorer rows, screen frames, and elements. A call is bracketed as it starts and ends, and failures can surface in the editor.
Activity marks help you locate work; they are not an audit log or proof that the result is correct. Re-read the changed structure and preview it.
The packaged desktop application must be running for its production MCP endpoint to exist.
Preview, screenshot, interaction, PDF, video, and other host-assisted operations depend on a healthy connection to the desktop host.
A project, file, node, or typed reference can be rejected when it is missing, stale, selected in another project, or incompatible with the requested field.
UI-only navigation state is not reproduced through project tools.
Developer-window tools used to test Misaki Studio's own interface are not part of the ordinary customer authoring workflow.
A successful tool response is not proof of visual quality, application compatibility, or a successful external deployment.
Connect to Misaki Studio, list the available projects, and let me choose one. After I choose, select it and read its full file tree. Do not create, update, move, export, publish, or delete anything.
This establishes the connection and project context without authorizing a mutation.