MCP Overview

The Model Context Protocol (MCP), open-sourced by Anthropic, has emerged as the industry standard for bridging AI assistants with external developer tools, private documents, and enterprise systems.

MuseCanto natively implements a remote MCP Server compliant with the latest spec (over Server-Sent Events / HTTP POST). This allows Cursor IDE, Claude Code CLI, ChatGPT, Grok, and Windsurf to search your knowledge base, query editorial topics, and draft content from the client you already use.

Connect Your AI Client

The connection URL is https://mcp.musecanto.com/mcp. Put that URL in your client's MCP configuration and authorize it in the browser the first time. Settings copies the same address.

Save your first content without setting up a library: choose organizationSlug, then call create_document or create_x_thread_draft without bookId. MuseCanto saves it to that workspace's Content inbox. An explicit standard library ID remains supported.

The inbox is created on the first successful save and counts toward the library quota. Saving a draft does not create a calendar plan or publish it. Articles return initial collaboration version: 0, location and a link; X drafts return updatedAt for subsequent edits.

Continue editing and prepare delivery

Read the entire article with get_document, then call update_document with its stable ID, baseRevision from the returned version, full revised Markdown and a new idempotencyKey. Pass an optional title (1–150 characters) to rename the article in the same call; get_document returns that title as name. Omit title to leave the name unchanged. An empty or oversized title is rejected and nothing is saved. On a collaborative article, version history records the new title. Restart a paged read if the version changes. On DOCUMENT_CONFLICT, read and reconcile the latest article before trying again; do not simply increase the version. Identical retries replay their result for 24 hours without creating another revision.

Try: “Read my Remote Collaboration article, add concrete examples to the second section, preserve the rest and its images, and update the original.” Edits use normal collaborative steps and reach open web editors. Markdown replaces the full body; use the web editor for special rich-text blocks that Markdown cannot preserve.

When deriving an X draft, optionally pass sourceDocumentId to create_x_thread_draft. This records the source without synchronizing the two documents. For X edits, use expectedUpdatedAt from get_x_thread and supply an idempotencyKey for safe retries.

prepare_publish returns a plan without publishing or scheduling anything. X requires an explicit connected account. A scheduled plan also needs an offset-bearing timestamp and an IANA time zone:

{
  "organizationSlug": "your-workspace",
  "documentId": "x-draft-id",
  "method": "x_scheduled",
  "xUserId": "account-id-from-list_x_accounts",
  "runAt": "2099-01-01T08:00:00+08:00",
  "timeZone": "Asia/Shanghai"
}

Replace the example time and verify the returned local time. ready: true only means local checks passed; executed and scheduled remain false. prepare_publish accepts only x_immediate or x_scheduled for an X thread. MCP does not publish long-form articles.

Use list_publish_tasks, then pass an exact returned task ID to get_publish_task to inspect status, next action and scheduled snapshot divergence. Editing a draft never updates its scheduled snapshot automatically. Platform drafts, confirmation steps and unknown outcomes are not public success. Verify unknown outcomes on the platform before retrying.

Approve an X publication for your agent

1. The agent reads updatedAt from get_x_thread, then calls request_publish_approval with organizationSlug, documentId, expectedUpdatedAt, method, xUserId and a fresh idempotencyKey. Scheduling also requires runAt and timeZone.
2. Open the returned reviewUrl. Review the text, images, quote/reply, account and time, then check the box and authorize. Requests expire 15 minutes after creation and bind one version to the initiating connection. The agent must never automate this confirmation on your behalf.
3. The agent checks get_publish_approval, then calls execute_publish with organizationSlug and approvalId. Repeated execution returns the same task.
4. Query the returned taskId with get_publish_task. accepted: true means a task was created. Immediate delivery runs on the next Cron tick; only execution results and publication receipts establish public success.

Revoke approval before execution. Edits, expiration or a disconnected AI require a fresh request. Once a task exists, cancel it in the publishing queue while it is still pending. The approval covers that one X thread. MCP does not publish long-form articles.

Reuse articles and writing preferences

Ask your AI to find and read your existing articles with search_documents and get_document. No separate material library is required. For an article variant, pass sourceDocumentId and the desired channel to create_document; use create_x_thread_draft for X. Inspect existing variants with list_content_variants. Versions remain independently editable and are never generated for every channel automatically.

Save or clear your preferences using “Writing preferences” on the content home. You can also explicitly ask the AI to use get_writing_preferences / update_writing_preferences; updates require expectedRevision, a new idempotencyKey and userProvided: true. Only text you supply or confirm is stored, scoped to your user and workspace. Chat history is not collected automatically.

Jump directly to the setup guide for your environment:

Standard 4-Step Connection Flow

https://mcp.musecanto.com/mcp

What each tool does

Permissions & Safety Guardrails

MuseCanto enforces strict permission tiers on MCP tools: