Documentation

MCP Server

Connect any MCP client — Cursor, VS Code, Claude Code, Codex, ChatGPT developer mode or your own — to the Context Link MCP server over Streamable HTTP with OAuth.

Context Link MCP Server

Context Link runs a hosted Model Context Protocol server. Any MCP-capable client can connect to it and get four tools for searching, answering from, and writing to your connected sources.

This is the same server that powers the two first-party integrations, so if you use Claude or ChatGPT you almost certainly want one of those guides instead — both wrap this endpoint in a two-click setup and neither needs anything on this page:

  • Claude — the custom connector is this MCP server. Follow the Claude custom connector guide; one setup covers Claude on the web, Desktop, Cowork and Claude Code.
  • ChatGPT — the official Context Link app is an MCP app talking to this same endpoint under the hood. Follow the ChatGPT app guide.

This page is for everything else: Cursor, VS Code, Claude Code on the command line, Codex, ChatGPT developer mode, or a client you're building yourself.

Quick reference

Property Value
Server URL https://www.context-link.ai/mcp
Transport Streamable HTTP (POST/GET/DELETE on the same URL)
Authentication OAuth 2.1 — authorization code + PKCE (S256), with Dynamic Client Registration
Scopes context:read, context:write, offline_access
Tools get_context, ask_question, save_memory, get_memory
Rate limit 1,000 requests per 10 seconds per IP

Use the www. host. The server is reachable at https://www.context-link.ai/mcp — that is the URL to put in every client config.

Prerequisites

  • A Context Link account with at least one connection set up (Notion, Google Drive, a website, email, and so on).
  • An MCP client that supports remote servers over HTTP. Clients that only speak stdio need a bridge — see stdio-only clients below.
  • ask_question requires a Pro plan. The other three tools work on any active plan.

The tools

Tool Scope What it does
get_context context:read Semantic search across everything you've connected. Returns the matching snippets plus the sources they came from. Takes query, and an optional mode to weight results.
ask_question context:read Runs the same retrieval, then has an LLM compose one concise paragraph grounded in your sources, with numbered citations. Takes query and optional mode. Pro plan only.
save_memory context:write Saves content back to Context Link as a Memory. Takes content (markdown or plain text, up to 100KB) and an optional namespace — a dashed name like brand-voice. Without a namespace you get a timestamped one.
get_memory context:read Retrieves a Memory you saved earlier. Takes namespace.

mode maps to the named weighting profiles you configure on your Connections page — pass something like customer-support and the results lean toward the connections you've weighted for that mode.

The four Context Link tools listed in Claude Code: Get Context, Ask Question, Save Memory and Get Memory


Setting it up

Every client wants the same two things: the server URL, and permission to open a browser window for the OAuth handshake. The steps below cover the common ones.

Claude Code

From your terminal:

claude mcp add --transport http context-link https://www.context-link.ai/mcp

Add --scope user to that command if you want the server available in every project rather than only the current one.

Then run /mcp inside Claude Code, choose the Context Link server and pick Authenticate. Claude Code opens a browser tab for you (and prints the URL in case it doesn't):

Claude Code authenticating with Context Link, showing the OAuth authorize URL it opens in the browser

Approve the consent screen in the browser — this is where you grant the read, write and offline access the tools need:

The Context Link consent screen asking whether to authorize Claude Code, listing Context:Read, Context:Write and Offline Access

Back in the terminal, /mcp should now show the server connected and authenticated:

Claude Code showing the Context Link MCP server connected, authenticated, and exposing 4 tools

Cursor

Add the server to ~/.cursor/mcp.json (global) or .cursor/mcp.json in your project:

{
  "mcpServers": {
    "context-link": {
      "url": "https://www.context-link.ai/mcp"
    }
  }
}

Reopen Settings → MCP, and Cursor will prompt you to sign in the first time it calls a tool.

VS Code (GitHub Copilot)

Add the server to .vscode/mcp.json in your workspace:

{
  "servers": {
    "context-link": {
      "type": "http",
      "url": "https://www.context-link.ai/mcp"
    }
  }
}

Start the server from the gutter action in that file, then approve the sign-in prompt.

Codex

Codex reads MCP servers from ~/.codex/config.toml:

[mcp_servers.context-link]
url = "https://www.context-link.ai/mcp"

If you'd rather use skills than MCP, Context Link ships those too — see Ways to Talk to Context Link.

ChatGPT (developer mode)

ChatGPT users are usually better served by the official Context Link app, which is two clicks and needs no configuration. If you specifically want to wire the raw MCP server up yourself, enable developer mode in ChatGPT's connector settings, create a new connector pointing at https://www.context-link.ai/mcp, and choose OAuth for authentication. Leave any "requires OIDC / userinfo" option off — Context Link issues opaque OAuth tokens and does not expose a userinfo endpoint.

stdio-only clients

Some clients still only launch MCP servers as local subprocesses. Bridge to the remote server with mcp-remote:

{
  "mcpServers": {
    "context-link": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://www.context-link.ai/mcp"]
    }
  }
}

The bridge handles the OAuth dance and opens a browser window on first use.


Troubleshooting

"Unable to reach Context Link" / connection fails immediately. Check the host — it's https://www.context-link.ai/mcp, with the www.

The client connects but every tool call returns 401. The client isn't attaching a token. Look for an "authenticate" or "sign in" action in the client's MCP panel; most only trigger the OAuth flow on demand rather than at connection time.

ask_question returns a pro_required reason. That tool is Pro-only. get_context gives you the underlying snippets on any plan.

Tools return no context. The token is bound to the Context Link user who approved the consent screen. If that user's account has no connections, or the content sits on another teammate's personal connection, there's nothing to search. Check the Connections page while signed in as the same user.

A tool call reports the allowance is exhausted. ask_question spends from your organisation's monthly LLM allowance. It resets at the start of each month.

Your client insists on OIDC. Turn it off. Context Link serves /.well-known/openid-configuration for discovery compatibility, but it isn't a full OpenID Provider — no ID tokens, no userinfo, no JWKS.