MCP 工具参考

Get the full content of a document section by document ID and section ID. Returns the section text and heading context; records a document-view audit event.

Edit a document, manage access, or upload content from a file.

ACTIONS - "create": title, [content], [folderId] — create with inline content. - "createFromTemplate": templateDocumentId, title, [folderId], [variables] — create from a template, substituting its declared variables. Locate templates with find; they live in templates folders and declare name, description, and variables in frontmatter. - "move": documentId, folderId — change a document's audience by moving it into that folder. A personal folder is owner-only, a tier folder is visible to its members; to make a document private, move it to the user's personal folder. - "edit": documentId, oldString, newString, [replaceAll] — targeted replacement. - "insertSection": documentId, position, markdown — insert blocks at {kind: beforeSection | afterSection, sectionPath} or {kind: endOfDocument}. Use heading paths from read_document. - "write": documentId, newContent — full INLINE replacement. - "share": documentId, email, role — grant access and email an invitation to an external recipient when needed. - "unshare": documentId, targetUserId — revoke access. - "presignUpload": contentLength + (documentId OR title [+folderId]). Returns { uploadId, uploadUrl, requiredHeaders, expiresAt }. - "finalizeUpload": uploadId. Idempotent. Returns { documentId, bytes, created }.

Upload flow (prefer when content already exists as a file or as bytes you generated — avoids re-emitting them as tool input): 1. wc -c file.md → byte count 2. presignUpload(contentLength, ...) → { uploadUrl } 3. PUT file to uploadUrl with requiredHeaders 4. finalizeUpload(uploadId) → { documentId }

Code blocks with language html render as live sandboxed previews (no network beyond them; vetted libs mermaid, d3, three, doom) and can chart named data from the same document: name a json/csv/tsv/yaml block on its fence (``json name=sales), a table with an HTML comment on the line above it (<!-- strata:name=sales -->), or use the reserved name frontmatter; the html block declares what it reads (``html uses="sales") and gets it inside the preview as strata.data.get('sales') — authoring guide: find(scope='strataDocumentation', slug='interactive-html'). Blog frontmatter controls a published post's design and metadata; edits report blogValidation findings. Guide: find(scope='strataDocumentation', slug='blog-design').

Export a presentation as standalone HTML or printed PDF. Returns a download URL valid for fifteen minutes. Presenter notes are removed from slide decks; opaque whole-deck HTML is rejected because note removal cannot be proven. PDF consumes the document owner's capture allowance; HTML is unmetered. PDF prints every slide in one headless page under a single sixty-second render deadline, with WebGL software-rendered and animation frozen to a still once a small render budget is spent, so a deck of many heavy scenes can time out where the HTML export succeeds.

Search, list, or discover documents, spaces, and folders across the workspace. Use this tool whenever you need to locate content before reading it — set intent to the kind of discovery (search, list, recent, related, info) and scope to where to look. List folders and their scopes (a document's audience comes from its folder) with intent='list', scope.type='folders'. Important: scope is a JSON object with a type field, not a string. Examples: {"scope":{"type":"all"}}, {"scope":{"type":"space","id":"space_..."}}, {"scope":{"type":"document","id":"doc_..."}}. Search uses hybrid keyword + semantic retrieval fused via Reciprocal Rank Fusion, so both short keywords and natural sentences work as queries. Use filters for status, owner, date ranges, includeShared, direction, sinceDays (for intent='recent'), and tokenBudget. Returns a uniform shape with items that carry id, title, snippet, score, and source attribution; follow up with read_document to fetch full content for any item you want to use. If you receive zero results, broaden the scope before reformulating the query — narrow scopes can return empty even when content exists.

Read company theme CSS and its verified source version for copying into a presentation.

Get the connection graph neighborhood around a document. Returns nodes (documents) and edges (connections) discovered by BFS traversal.

Optional parameters: - depth: BFS depth (default 2, max 5) - maxNodes: Maximum nodes to return (default 50, max 200)

Retrieve an image from a document as base64-encoded data. Use when you need to analyze, describe, or understand an image referenced in a document.

The imageId is found in document content as [Image: alt text (imageId: <id>)] markers or ![alt](strata://image/<id>) in markdown.

Returns the image as inline base64 PNG data that can be viewed directly.

List company presentation themes available to copy.

Read a document's content by ID.

Returns rendered Markdown. Use before editing. Records a document-view audit event. Large documents automatically return a structural outline instead of full content (default budget: 10000 tokens). Use the sections parameter with heading paths from the outline to read specific sections.

Workflow: read_document(id) → outline (if large) → read_document(id, sections: ["Heading > Subheading"]) → section content.

Render presentation slides and return layout, contrast, binding, structure, and theme findings. Always executes checks and stores captures. Costs one capture per slide per viewport, with no repeat charge for the same render key this cycle. Checks at most two slides at one viewport per call. Repeat with next as continuation and unchanged slides/viewports while status is more; merge findings and checked pairs. Omitted selection covers all slides at both viewports.

Query who is currently viewing or editing a document. Returns a list of collaborators with their entity type, display name, active section, status, and presence color. Useful for understanding who else is working on the document before making edits.

Manage comments on a document. Create/reply can invoke mentioned agents and their connected tools.

Actions: - "create": Create a new comment. Requires "body". Optionally provide "quotedText" to anchor the comment to specific text, and "headingContext" for disambiguation. - "reply": Reply to an existing thread. Requires "parentCommentId" and "body". - "resolve": Resolve a comment thread. Requires "commentId". - "reopen": Reopen a resolved thread. Requires "commentId". - "list": List comments on the document. Optionally filter by "status" ("active" or "resolved").

Manage suggestions on a document.

Actions: - "propose": Create a new suggestion. Refresh proposals carry baseVersion, sourceRef {documentId, sourceVersion}, and anchor {sectionPath, blockOrdinal, offset}. Requires "originalText" (exact text to replace) and "replacementText" (proposed replacement markdown). Optionally provide "headingContext" for disambiguation, "severity" (info/warning/error), "confidence" (0.0–1.0), and "rationale". - "accept": Accept a pending suggestion (applies the edit). Requires "suggestionId". - "reject": Reject a pending suggestion (no edit applied). Requires "suggestionId". - "withdraw": Withdraw your own pending suggestion. Requires "suggestionId". - "list": List pending suggestions on the document.

Check the publish status of a document. Returns the current state (draft/published), version number, and whether there are unpublished changes (content hash comparison).

Publish a document to your public profile. Creates a frozen snapshot of the current content. If the content hasn't changed since last publish, returns the existing version. Every successful call records a publication audit event.

Side effects: Creates a PublishedVersion record and copies Yjs state to a versioned S3 key. Your creator profile is provisioned at signup, so no profile is created as a side effect of publishing.

Unpublish a document, removing it from public view. The document returns to draft state and public URLs return 404. Published version records are retained for potential rollback.

Check the status of an agent task.

Returns the current status, tool call count, and output (if completed) or error (if failed). Poll this after invoke_agent to track completion.

Possible statuses: pending, running, completed, failed, cancelled, timedOut.

Invoke a platform agent to perform a task.

Specify the agent by ID or by name (mutually exclusive). Optionally provide a target document and a natural-language instruction.

Returns a task ID — use get_agent_status to poll for status and output. The agent executes asynchronously; this tool returns immediately after the task is created.