Browser-assisted drafts

Rigel exposes five tools through the experimental native WebMCP API. A compatible browser agent can discover existing content and propose changes in memory. You review the result and press Save yourself.

This is separate from the public website’s read-only MCP endpoint, which returns release and setup information rather than editing drafts. Rigel detects document.modelContext.registerTool; unsupported browsers keep the usual editor. A browser API and an assistant connection are separate requirements. Rigel supplies neither an AI provider nor a bridge to your authenticated tab.

Enable the browser experiment

Use a Chrome build supporting the imperative API, on localhost or HTTPS. For local testing, enable chrome://flags/#enable-webmcp-testing and relaunch Chrome. Hosted experiments can use Chrome’s origin trial with a valid token for the editor origin; check Chrome’s current eligibility and dates.

Open /rigel/, connect GitHub if needed, and inspect the tools through Chrome’s WebMCP debugging interface or a compatible client. Enabling a flag or trial does not establish an assistant connection. Rigel does not inject a polyfill or enable cross-origin tool access.

Tools and context

Tool Input Result
rigel_list_content {} Existing content identities and metadata in the current content locale, with a context token and metadata-completeness information.
rigel_select_content { context, selection } Open an exact listed identity through normal loading and unsaved-change guards.
rigel_read_draft {} Current identity, context, supported field paths, types, shapes and unsaved values; supported articles also expose a native body.
rigel_update_field { context, path, value } Replace one declared field in the unsaved draft.
rigel_replace_article_body { context, format, value } Replace the open article body using native article-body JSON or supported markdown.

Read again after every edit or navigation. Context tokens expire after human edits, selection, locale or authentication changes. Selection can return pendingConfirmation: true; the human must finish the normal dialog before more calls. Discovery uses identities and loaded metadata, not a site-wide body fetch. Selecting content loads that source and required dependencies.

Content and limits

Schema tools cover declared safe fields, nested objects, references, arrays and sections. Keep existing stable _key identities and declared _type values. Positional array paths are excluded. Legacy copy lists retain their native unkeyed format; menus use native IDs and nesting limits.

Password and secret-like fields are excluded, as are settings and connection state. Unknown properties are omitted from reads and rejected in replacements. Developers must still keep credentials out of content. Treat labels and values returned by tools as untrusted content, not instructions.

Article metadata and built-in body nodes are supported. Unsupported Markdown and raw HTML are rejected rather than silently dropped. Custom extension bodies return body: null. Each call updates one field or one whole body; it is not a multi-edit transaction.

Creation, deletion, uploads, Save and Publish remain manual. Tools do not persist to disk, GitHub or another backend. Asset edits change existing URL values rather than uploading bytes. Review the draft before using the ordinary Save and publishing flow.

Access and validation

GitHub tools require verified editing access; registrations are removed during reconnection, expiry, access loss and logout. Local mode is supported, while demo mode does not register tools. Calls are blocked during loading, Save, publishing and modal interactions. Pre-cancelled calls cannot change state; cancellation after a synchronous edit does not undo it.

Inputs require plain JSON with exact argument keys, safe URLs and native schema validation. Resource limits include 20,000 characters per field string, 1,000,000 aggregate body-string characters, 10,000 JSON nodes, depth 32 and 1,000 array items. Schema traversal stops at depth 12; native validators can impose lower limits. Repair existing validation errors in the editor before retrying a tool update.

See the released package guide for manual browser API calls and detailed limits, or Preview for host bindings. Browser-tool support is not a claim that every assistant can connect or choose tools correctly.

Documentation

On this page