Skip to content
Join the waitlist

Your agent, on the canvas

The Builder Studio MCP server gives your agent the canvas you work on, through 100 tools. It reads and writes the same pages you do, and publishes them.

How does an agent see your canvas?

canvas_outline reads a canvas as one line per layer, the cheapest look at its structure. canvas_read_html returns a page or a section as HTML and CSS.

canvas_screenshot renders a layer as an image, or as a sheet of its desktop, tablet and mobile widths, so the agent can look at what it made before it says it is done.

How does it change things without overwriting yours?

canvas_write_html writes new pages and sections. canvas_apply_patch changes text and styles or deletes layers, and canvas_place_nodes moves and copies them.

A write carries a fence, a mark of what the agent last read. If you or another agent changed those layers since, the write is refused with VERSION_CONFLICT and the agent reads again, so nobody's edit is lost.

Can it publish the site?

Yes. site_check_publish lists what would stop a publish, and site_publish takes the site live. The CMS tools create collections and write their items, up to 25 in one call.

An agent can also import a WordPress site, export the site as an Astro project, or open a draft pull request in a GitHub repository you connected.

Can it work with Builder?

Yes. builder_chat_send hands a request to Builder, the agent inside the product, and builder_chat_get reads what it did. Your coding agent can ask Builder for an image or a layout the way you would.

The tool sets

100 tools, grouped as the server groups them. 55 are experimental and say so at the start of their description.

  • Canvas

    39 tools. Read a canvas's structure, write pages and sections as HTML, patch and place layers, render screenshots, answer comments and draw diagrams.

  • Sites and CMS

    16 tools. Check and publish a site, roll it back, write collections and items, import a WordPress site, export an Astro project and read analytics.

  • Assets

    14 tools. Upload media, import it from a link, and list and place it on a canvas.

  • Design system

    7 tools. Read and change variables and styles, and import or export the system.

  • Prototypes

    6 tools. Make a working prototype in React or plain HTML that runs live on a canvas, change its files, try variations side by side and read who changed what.

  • GitHub

    6 tools. List repositories, read and sync files, and open a draft pull request.

  • Skills

    3 tools. List, read and save the instructions your workspace keeps for agents.

  • Builder

    3 tools. Hand Builder, the agent inside the product, a request and read its reply.

  • Workspace

    3 tools. Find a tool by its job, read the workspace and list its API keys.

  • Google Docs

    2 tools. Import a Google Doc onto a canvas, or export a document to Google Docs.

  • Billing

    1 tool. Read the workspace's AI balance before a run that spends it.

Sign your agent in

An agent such as Claude Code connects with OAuth: it opens a sign-in page, you choose the workspace it may use, and its token works for that workspace alone. Scripts and servers use a workspace API key, created in Workspace settings › API & agents › Access keys.

  • Every tool declares the scopes it needs, and tools/list shows only the tools a grant can call.
  • A scope to write includes reading; deleting is never included in writing.
  • Rate limits apply per key and per workspace, and a refused call says how long to wait.

Add Builder Studio to Claude Code

Terminal
claude mcp add-json --scope user builderstudio '{"type":"http","url":"https://builder.studio/api/mcp","headers":{"Mcp-Readback":"text"},"oauth":{"scopes":"canvases:read canvases:write assets:read assets:write skills:read offline_access"}}'
The command registers the server with OAuth, so Claude Code signs in with your Builder Studio account instead of an API key.

Questions

Which agents can connect?

Any agent that speaks MCP. The setup page has steps for ChatGPT, Claude, Claude Code, Codex, Cursor, OpenCode and Grok. Connect your agent.

Which tools does my agent see first?

The default address lists 22 tools for everyday canvas work. tool_search finds the others by job and returns their inputs, and the agent can call any of them by name.

What does an experimental tool mean?

Its name, inputs or behaviour may still change. Each experimental tool says so at the start of its description, so your agent knows before it relies on one.

Can a key reach every workspace I belong to?

No. An API key belongs to one workspace and can be narrowed to some of its sites, and an OAuth token works only in the workspace you chose when you signed in.

Give your agent the canvas

Pick your agent on the setup page and sign in with your Builder Studio account.