Docs · Reference

MCP server reference

Redtrail is a remote MCP server that lets an agent read and edit the Word document its user has open on SharePoint or OneDrive, in place, as tracked changes, inside the live co-authoring session. This page is the reference an MCP client or directory needs: endpoint, sign-in, scopes, tools and connect snippets.

Before you connect

The server edits through the Redtrail add-in in your own Word session. Your Microsoft 365 tenant must have the add-in deployed (or sideloaded for a trial), and you must have a document open with the task pane connected. Otherwise every tool returns an empty document list. See the quickstart.
Sign-in is your Microsoft 365 account, through the server's own OAuth 2.1 flow. No Copilot license is involved. Free tier: one document per session.

Endpoint and sign-in

Endpoint	`https://redtrail.ai/mcp`
Transport	Streamable HTTP (MCP 2025-06-18); JSON responses, SSE for long calls
Authentication	OAuth 2.1 with PKCE. The server is its own authorization server at `https://redtrail.ai`; Microsoft Entra is the upstream identity provider, so the consent screen is the Microsoft account picker.
Client registration	Dynamic client registration at `/oauth/register`; redirect URIs must be https, loopback http, or a private-use scheme. Rate limited to five registrations per minute per address.
Discovery	`/.well-known/oauth-protected-resource/mcp` and `/.well-known/oauth-authorization-server`; an unauthenticated call to the endpoint returns 401 with the resource-metadata challenge.
Scopes	`docs.read` (reads), `docs.write` (writes, comments, accept and reject), `history.read` (history queries). Unknown scopes are refused with `invalid_scope`; a read-only grant gets `insufficient_scope` on a write.
Tokens	One-hour access tokens, rotating refresh tokens with reuse detection. Grants are named after the device and can be revoked from the task pane or by an admin.

Connect from your client

Claude Code

claude plugin install Redtrail
# or, without the plugin's skill and etiquette rules:
claude mcp add --transport http Redtrail https://redtrail.ai/mcp

Claude.ai, Claude Desktop and Cowork

Settings → Connectors → Add custom connector, URL https://redtrail.ai/mcp. The first tool call opens the Microsoft sign-in.

Cursor

{
  "mcpServers": {
    "Redtrail": { "url": "https://redtrail.ai/mcp" }
  }
}

Visual Studio Code

{
  "servers": {
    "Redtrail": { "type": "http", "url": "https://redtrail.ai/mcp" }
  }
}

Codex, Copilot Studio and other clients

Add a remote (streamable HTTP) MCP server with the endpoint URL and OAuth. Any client that implements the MCP authorization flow, including dynamic client registration, connects without further setup. Tenants on Microsoft Agent 365 can register the server as a bring-your-own MCP server; see the quickstart.

Tools

Eighteen tools. Reads are answered immediately. Writes are applied by Word in the user's session as tracked changes, either at once (auto-apply, only while tracking is on) or after the user clicks Apply in the task pane; the call returns status: pending after a short hold rather than blocking.

Tool	Scope	Purpose
`doc_list`	read	Documents this user has open with the add-in connected: id, title, location, tracking and auto-apply state
`doc_outline`	read	Headings with paths, content-control tags, per-section paragraph counts, comment and change counts
`doc_read`	read	Text of a scoped region, by anchor, with inline anchors
`doc_search`	read	Search hits with anchors (query, occurrence, context hash) and surrounding text
`doc_changes`	read	Tracked changes: id, type, author, date, text, anchor
`doc_queue_status`	read	What actually landed: waves, per-change status, conflicts
`doc_replace`	write	Replace anchored text as a tracked change; carries the expected before-text
`doc_insert`	write	Insert before, after or inside an anchor as a tracked change
`doc_comment`	write	Add a comment anchored to a range
`doc_reply`	write	Reply to a comment thread
`doc_resolve`	write	Resolve a comment thread
`doc_accept`	write	Accept a tracked change by id
`doc_reject`	write	Reject a tracked change by id
`doc_history_log`	history	Agent change history for a document: who, which agent, when, operation, anchor, verification
`doc_history_show`	history	One history event in full
`doc_history_blame`	history	Which agent events last touched an anchor, across the fork family
`doc_history_revert`	write	Apply the inverse of one event as a new tracked change
`doc_history_replay`	write	Replay events onto another document, usually a fork

Anchors and verification

Anchors are never paragraph indices. A write targets a content-control tag, a heading path, a search hit with an occurrence index and context hash, a comment id or a tracked-change id. Every write re-resolves its anchor when it runs and carries the text the agent expects to replace. If the current text differs, the write is reported as a conflict with what was expected and what is there now, and nothing is applied. A missing anchor is not_found. Both are recorded in the history as failed events.

What the server expects from an agent

Read the outline, then only the section you need.
Comment before rewriting anything longer than a sentence.
Keep tracked changes on; do not ask the user to turn them off.
After a batch of writes, call doc_queue_status and check what applied before continuing.

The Claude Code plugin ships these as a skill. Other clients get the same rules from the tool descriptions.

Data handling

The server is a content-blind relay: it routes on envelope headers and never stores document text. Tool calls are executed by Word in the user's session. The optional change history stores events, not copies of the document, encrypted under a key the tenant can hold. Details on the security page; terms and privacy are linked in the footer.