Skip to content
CollabHub.ai
Sign inGet started

Connect an AI client

CollabHub speaks MCP (Model Context Protocol), so the AI tools you already use, such as Claude Desktop, Claude Code, Cursor, ChatGPT and VS Code, can work with your lab directly, always under your own permissions.

Connect in two minutes

Every client points at the same remote server:

Remote server URL
https://mcp.collabhub.ai/mcp
  1. Add the URL to your client. The exact setup for each one is below, and on Connections settings (sign-in required) once you’re signed in. Claude Desktop adds it as a custom connector under Settings → Connectors, Claude Code is one terminal command, and Cursor and VS Code take a small mcp.json entry. ChatGPT connects the same way if your ChatGPT plan includes custom connectors.
  2. The first time your client reaches CollabHub, your browser opens a one-time sign-in. There is no token to copy and no secret in your config file.
  3. Review the consent screen. It names the app, the workspaces it can reach and each kind of access it asks for, with read and write access listed separately. Nothing is shared until you approve.
  4. Ask your client to search your workspaces to confirm the connection.

Setup for each client

Claude DesktopAdd CollabHub as a connector in the desktop app.
Steps
Add CollabHub to Claude Desktop as a custom connector (no config file to edit):

1. Claude Desktop → Settings → Connectors → Add custom connector.
2. Name:  CollabHub
3. Remote MCP server URL:  https://mcp.collabhub.ai/mcp
4. Add it, then sign in with your CollabHub account when the browser opens (OAuth).
5. Ask Claude to search your workspaces to confirm the connection.
  1. Settings → Connectors → Add custom connector.
  2. Paste the remote URL: https://mcp.collabhub.ai/mcp
  3. Approve the one-time OAuth sign-in — there is no token to copy.

Prefer editing claude_desktop_config.json? Claude Desktop can't take a bare remote URL there (it accepts stdio servers only), so use the mcp-remote bridge instead of a raw url entry: {"mcpServers":{"collabhub":{"command":"npx","args":["-y","mcp-remote","https://mcp.collabhub.ai/mcp"]}}}

Claude CodeOne command in your terminal.
Terminal
claude mcp add --transport http collabhub https://mcp.collabhub.ai/mcp
  1. Run the command in your terminal (inside any project).
  2. On the first CollabHub tool call, Claude Code opens your browser to sign in with CollabHub.
  3. Verify it registered with `claude mcp list` (look for "collabhub").
CursorDrop a project or global mcp.json.
Config
{
  "mcpServers": {
    "collabhub": {
      "url": "https://mcp.collabhub.ai/mcp"
    }
  }
}
  1. Create or open .cursor/mcp.json in your project (or ~/.cursor/mcp.json for every project).
  2. Merge in the snippet and save.
  3. Open Cursor → Settings → MCP; complete the CollabHub sign-in when Cursor prompts you.
VS CodeAdd an http server to .vscode/mcp.json.
Config
{
  "servers": {
    "collabhub": {
      "type": "http",
      "url": "https://mcp.collabhub.ai/mcp"
    }
  }
}
  1. Create .vscode/mcp.json in your workspace (or add to your user mcp.json).
  2. Merge in the snippet and save.
  3. Run "MCP: List Servers" from the Command Palette, start CollabHub, and sign in when prompted.

What a connected client can do

With read access, a client can:

  • Search across your workspaces, list their datasets, and query the knowledge graph.
  • Read notebook entries, datasets, protocols, and experiments. Entries arrive as clean Markdown — rich text, tables, code, and equations included — so models work with your real records, not screenshots.
  • Look up file details (name, type, size, scan status). File contents stay behind the app’s own download checks and are never streamed to a client.
  • Browse and read items shared to the public library.

If you also grant write access, it can:

  • Log experiment observations in batches of up to 100 in a single call.
  • Upload files and datasets into a workspace.

Every action runs as you. A client can never see or do more than your own role in each workspace allows, and every call is recorded in the audit trail.

Consent, control and revoke

  • Consent first. Access only exists after you approve it on the consent screen, and only with the access listed there.
  • See every connection. Connections settings (sign-in required) lists each connected app with what you granted and when.
  • Revoke instantly. One click disconnects an app, and every credential it holds stops working immediately.

Advanced: tokens and self-hosting

  • Personal access tokens. Create them on the same Connections page, each with its own scopes, an optional expiry and an optional list of workspaces it is limited to.
  • A token is shown once, stored only as a hash, and can be revoked at any time.
  • Self-hosting. Operators who run their own CollabHub can run the MCP server locally over stdio. The Connections page includes that setup too, marked as operator-only.