Connect AI agents to Kord

The server lives at a single endpoint:

https://mcp.withkord.com/mcp

The first time your agent calls Kord, your MCP client opens a browser tab where you sign in with the Google or Microsoft account you already use for Kord and click Allow. That's the whole setup - the agent then acts with your permissions and nothing more.

What a connected agent can do

Read tools cover the full working context: list your teams and repositories, browse the file tree, read a file's version history, and inspect review sessions - files, reviewers, comments, markups - plus change requests and issues.

• Open a review session with changed files. An agent that edited documents on your machine can stage the files and open a review session in Kord, ready for your reviewers - instead of leaving changed files in a folder for you to upload by hand.
• Add files to an open review session - new files, or new revisions of files already under review.
• Link a cloud folder as a new repository - the agent equivalent of "Track folder from cloud", using a cloud connection you've already set up inside Kord.
• Answer questions about your project - "which review sessions are still waiting on me?", "what changed in the last revision of the vessel datasheet?", "which change requests are still open on this repository?"

Claude Code

Add the server from your terminal:

claude mcp add --transport http kord https://mcp.withkord.com/mcp

Then run /mcp inside Claude Code and pick kord to complete the browser sign-in. From then on Claude Code can browse your repositories and open review sessions directly.

Claude.ai and Claude Cowork

Kord connects to Claude.ai as a custom connector, which also makes it available in Claude Cowork sessions:

1. Open Settings → Connectors in Claude.ai.
2. Click Add custom connector and paste https://mcp.withkord.com/mcp.
3. Click Connect and complete the browser sign-in with your Kord account.

On a Claude Team or Enterprise plan, an organization owner adds the connector once under the organization's Connectors settings, and each member just clicks Connect.

ChatGPT Work

Kord connects to ChatGPT as a custom connector, so ChatGPT Work can read your document sets and open review sessions the same way:

1. Open Settings → Connectors in ChatGPT.
2. Add a custom connector pointing at https://mcp.withkord.com/mcp.
3. Complete the browser sign-in with your Kord account. On a workspace plan, an admin adds the connector once and each member connects with their own Kord login.

Cursor, Windsurf, and other MCP clients

Any client that speaks MCP over HTTP works. In a JSON-configured client, add a server entry pointing at the URL (the exact key names vary slightly by client):

{
  "mcpServers": {
    "kord": { "type": "http", "url": "https://mcp.withkord.com/mcp" }
  }
}

Authentication and permissions

Authentication is OAuth 2.1 - the same standard your MCP client already uses for other connectors. There are no personal access tokens and no shared API keys: the sign-in mints a session tied to your Kord account, and every tool call runs under your exact access. An agent can only see the teams, repositories, and review sessions you can see, and its actions show up in the activity log under your name.

You can review and revoke connected agents at any time from your Kord settings.

For a client wiring this up by hand: an unauthenticated request to the endpoint gets a 401 whose WWW-Authenticateheader names the server's protected-resource metadata at mcp.withkord.com/.well-known/oauth-protected-resource. That document (RFC 9728) lists the authorization server and the scopes it accepts - the OpenID identity scopes openid, email and profile; what an agent can do is bounded by your Kord access, not by scope. The authorization server's own metadata (RFC 8414) and OpenID configuration are one hop from withkord.com/.well-known/oauth-authorization-server and withkord.com/.well-known/openid-configuration.

The Kord site MCP server (no account needed)

Everything above is the product server, which needs your Kord sign-in. withkord.com also runs a second, entirely separate MCP server that is public and unauthenticated. It serves this website rather than your documents, and it can mint a diff link:

https://withkord.com/api/mcp

Every page of the site is exposed as an MCP resource with text/markdown content, at its own canonical URL, plus four tools:

• list_pages - every page on the site with its path, title and description.
• search_site - full-text search across all of them, returning excerpts and the path to read next.
• read_page - one page as markdown, by path or by full URL.
• create_diff_link- the one tool that acts: two files the user has, as text or base64, and back comes a Kord link that shows their diff to anyone who opens it, for seven days. No account on either end. This is what "diff these two files with Kord" resolves to in a chat assistant.

Point an agent at it to answer questions about Kord - what it does, who it is for, how it compares to Bluebeam, PLM or a CDE, which storage it integrates with - without scraping HTML, and to turn two files into a diff link. It is stateless and knows nothing about your account, so there is nothing to sign in to and nothing to revoke; a diff link it mints is an ordinary anonymous Kord link, rate-limited and budgeted on Kord's side. The same endpoint answers at withkord.com/mcp for a client whose Accept header marks it as MCP rather than a browser, which is why this page and the server can share a URL.

The same pages are available two other ways, no MCP client required: send Accept: text/markdown to any page URL to get that page as markdown, and read withkord.com/llms.txt for a machine-readable index of the whole site.

Diff two files with no account at all

Not everything needs the product server. When a user hands an agent two files - an original and a revised copy - and asks what changed, the agent gets a link to their diff with one HTTP call and no sign-in:

curl -F before=@spec-rev-a.pdf -F after=@spec-rev-b.pdf \
  https://work.withkord.com/api/diff/share

The JSON response carries a url: a page that renders the comparison for anyone who opens it, in the same viewer as the free diff tool - spreadsheets cell by cell, PDFs page by page, Word as a redline, 3D models in the browser. The link lasts seven days by default (up to 30 with ttlDays), the two files are kept only until then, and nobody can list or revoke an anonymous link - which is why it is short-lived. Files over about 4 MB together go through /api/diff/upload-url first. The full contract is in Kord's OpenAPI document, and work.withkord.com/llms.txt says the same for an agent reading the app host directly.

A chat assistant cannot make that call itself: claude.ai and ChatGPT run their code with no internet access, so the files never leave the sandbox. Add the site MCP server below as a custom connector instead - no sign-in - and ask. Its create_diff_linktool makes the same call from the platform's own servers and hands back the link. Files up to 1.0 MB each fit a tool call; bigger ones go to the diff page.

Requirements

You need a Kord account, the connector signs in with it. If your team isn't on Kord yet, chat with us and we'll get you set up.