Skip to main content
Version: Current

Use Scribe with AI agents and WebMCP

WebMCP lets a compatible AI agent work with the Scribe project you already have open. Scribe publishes typed site tools for intentional editor operations—such as inspecting scenes, editing text, arranging objects, adding keyframes, or preparing an export—without giving the agent raw application state or relying on button clicks.

What stays the same​

  • Agent actions update the real, visible project. There is no hidden second editor.
  • Normal authentication, project permissions, plan limits, media limits, and export limits still apply.
  • Reversible local edits use Scribe history, so Undo remains available.
  • Locked objects stay locked until you explicitly unlock them.
  • Scribe does not host an LLM for this feature. Intelligence comes from your compatible browser agent.

What an agent can do​

The Desktop editor exposes a discoverable semantic catalog across projects, scenes, objects, selection, layout, text, equations, assets, vector tracing, boards and drawing materials, Smart Connectors, Timeline/keyframes/curves, Draw Steps, camera, path following, audio, video, screen recording, templates and AI JSON, persistence, export, sharing, account capabilities, plugins, and useful editor views.

The complete catalog remains searchable, but Scribe advertises a smaller client-safe working set at any one time. Core discovery, project/scene context, batching, history, selection, and playback tools stay available. A compatible agent can search the full catalog and activate the exact domain or tools needed for the current task; the browser's standard tool-change signal refreshes the list without clicks or page reloads.

Availability is contextual. For example, a search can find video-trim capabilities while explaining that they are unavailable when the project has no video object. Ask the agent to inspect capabilities and project state before making a complex change.

Shapes catalog versus primitive shapes​

Scribe's Shapes tab contains a public catalog of editable SVG presets. A compatible agent can search that same catalog with the shape asset type, inspect stable preset metadata, and insert a selected result as a native drawing. This is separate from creating primitive rectangles, circles, or polygons. If you want a catalog arrow, star, or speech bubble, tell the agent to use the Shapes catalog rather than Drawings or primitive shape creation.

Human-controlled steps​

WebMCP does not bypass protected browser or account surfaces. You still choose or approve:

  • local files;
  • the screen, window, or tab used for screen recording;
  • microphone permission;
  • credentials, OTPs, and reauthentication;
  • purchases and billing approval;
  • destructive confirmations such as account deletion; and
  • browser/native download or share destinations when the platform requires it.

The agent can prepare the correct Scribe workflow, but the browser or operating system owns the protected choice.

Use a compatible client​

  1. Open the Scribe Desktop web editor and the project you want to change.
  2. Open the page with a client that supports top-level imperative WebMCP tools.
  3. Inspect the site's available tools in the client before granting consequential work. For a specialized task, ask it to search Scribe capabilities and activate the relevant domain.
  4. Start with a read-only request, such as “Inspect the current project and list locked objects.”
  5. Ask for a small reversible change and verify it on the canvas.
  6. Use Scribe Undo if you do not want the change.

OpenAI currently calls its WebMCP implementation Site tools in the ChatGPT desktop app's built-in browser. Availability depends on client version, rollout, workspace, and model. Merely having a browser extension installed does not prove that extension supports WebMCP; a compatible client must actually discover and invoke Scribe's registered tools.

If the browser does not implement document.modelContext, Scribe behaves normally and no site tools are registered.

Developer testing​

In local development, Chrome's current test implementation may require:

chrome://flags/#enable-webmcp-testing

Relaunch Chrome after changing the flag, open the editor, and check:

await document.modelContext.getTools()

getTools() returns the current bounded working set, not the entire canonical catalog. To inspect and activate a tool that is not currently advertised, execute scribe_search_capabilities, then scribe_activate_capabilities; the normal toolchange event updates the browser's list. Scribe keeps the simultaneous advertised count below current client limits while preserving full semantic reachability.

Use /dev/webmcp for Scribe's sanitized support status, origin-isolation state, platform profile, tool count, schemas, availability, and native executeTool() smoke tests. This route is development-only.

Production Chrome may require an origin-trial token while the standard is experimental. Deployment operators can provide WEBMCP_ORIGIN_TRIAL_TOKEN; Scribe does not include a personal token in source control.

Security and privacy​

Scribe tools are same-origin by default and are not exposed to wildcard origins. Public share and embed routes do not receive editor tools. Project/template text and imported metadata are marked as untrusted content for compatible agents. Tool results exclude cookies, credentials, tokens, private signed upload details, media bytes, and private Zustand or React state.

This page is the public behavior and compatibility contract. Private repository structure and internal implementation modules are not public integration surfaces.

Last updated on