Working with an AI assistant (MCP)

The Traceable MCP connector lets an AI assistant such as Claude or ChatGPT work in your account the way you do: browse projects, read documents, draft and edit content, build ID grids, create trace links and search the Reference Library. It acts as you, changes drafts only, and leaves publishing and signing to people.

This is the short version. Every tool and limit is described in the MCP connector reference.

Two rules that govern everything

1. It acts as you. The assistant signs in with your account through a consent screen and reaches exactly what you can reach. If you are not a member of a project, neither is the assistant. If your role is read-only, so is the assistant.
2. It only touches drafts. Publishing, review and e-signature stay human actions in the app. Every change is recorded in the audit trail against your name.

Before you connect: two switches

Both of these must be on. If either is off, the connection is refused.

• For the org: an Admin ticks Allow external AI agents to connect (MCP) in Org settings > AI agents, where the connection URL to paste into your assistant is also shown. It is on by default. Turning it off stops every member's connected assistants straight away.
• For you: an Admin goes to Org settings > Members, selects Edit beside you, and switches on Enable MCP access. It is off by default.

If you belong to several orgs, every one must allow external agents, because the connection belongs to you, not to one org.

If you sign in before you have been given access, Traceable tells you which switch is off. When it is your own access, select Ask an administrator for access to email your org's Admins.

Connecting

The server address is:

https://app.traceabledocs.com/api/mcp/mcp

These five clients have been connected and used end to end:

• claude.ai (web, desktop and mobile): open Settings > Connectors, choose Add custom connector, paste the address and save.
• Claude Code: run claude mcp add --transport http traceable https://app.traceabledocs.com/api/mcp/mcp, then follow the sign-in prompt the first time a Traceable tool is used.
• ChatGPT: turn on developer mode, then add the address as a custom connector.
• Cursor: add the address in its MCP settings as an HTTP server.
• Google Antigravity: in the Agent panel, open ⋯ > MCP Servers > Manage MCP Servers > View raw config, add the address as serverUrl under the name traceable, save and choose Refresh. Its changes show as made by MCP CLI Proxy, the name it connects under.

Other clients that accept a custom server address, such as VS Code and Gemini CLI, should also work but have not yet been verified. A client that cannot talk to an HTTP server directly can use npx mcp-remote with the same address.

The consent screen

Each client sends you to Traceable to sign in, then to a screen headed Authorize followed by the app's name. It asks for two permissions:

• Read your projects and documents
• Create and edit document content

Under About this application, check Sends you back to. It is the one detail Traceable has verified, so if it is not an address you expect, select Deny. Otherwise select Approve.

What it can do

• Explore: list orgs, projects and templates, map a project's documents and trace links, and read a document.
• Author: create a document, optionally from a template, and insert, replace or delete content.
• ID grids: add and rename columns, change a column's type, and set cells, keeping every row's trace links.
• Traceability: create and remove trace links, rename a Trace ID without breaking its links, and change a document's ID prefix.
• Reference Library: search and read items you can access, and cite them in a document.
• Skills: list and read your organisation's skills, the reusable instructions written in Skill rows (see Rows and the editor), and follow the one that fits the task.

Name things precisely, using document numbers, Trace IDs and REF codes. Prefer targeted edits, such as "update FR-012", to rewriting a whole section, which deletes its rows and breaks links to them.

What it cannot do

• Publish, review, approve or sign.
• Create projects, orgs or users.
• Add images to document content.
• Delete a live document. Only an archived document can be deleted, and that is permanent.
• Read PDFs, PowerPoint files or images in the Reference Library.

What happens to your documents

A change to a published document opens a new draft first. The published version stays in the document's history, and the assistant's edits get their own line on the Versions timeline.

Each project has a Require AI change review setting, which an Admin manages in Org settings > Projects. It is on by default. While it is on, every row an assistant writes waits under Pending on the Agent changes tab of the document's Review panel, for a contributor or admin to Acknowledge. A changed row also carries an agent marker that opens it. Until its agent changes are acknowledged, the document cannot be published and an Approver cannot sign it.

Deleted rows can be recovered, and links to them are marked broken rather than disappearing.

Skill governance

On Regulated and Enterprise plans, two controls govern the skills an assistant works to. Where your organisation is under controlled releases, an Admin first turns on Skill governance in Org settings > Releases.

• Agents use published skills only: an Admin ticks this under Skills in Org settings > AI agents. While it is on, an assistant is never given a draft skill, however it asks, so it works only to methods that have been reviewed and published. Turning it off is always allowed.
• Skills on each change: when an assistant tells Traceable which skills it followed for a change, the change records them by name, document number and version. They are shown on the change's card under Agent changes and in the agent change window, so a reviewer can see which method shaped it. The record keeps saying what it said when the change was made, even after the skill is republished.

Managing access

Account > Security, under Connected apps, lists every app you have authorised. Select Revoke to disconnect one: it can no longer obtain access, and access it already holds stops working within the hour. See Your account.

An Admin can revoke a member's connection under Active connections in that member's settings, or switch off Enable MCP access to cut them off at once. See Members, roles and seats.

If it stops working

The connector is served by the Traceable app itself, so it drops briefly whenever a new version is deployed. Reconnect, or ask the assistant to try again.