Skip to main content
Profile MCP settings showing the MCP endpoint, create token action, and masked token list
Use MCP tokens to let external AI clients call Atomic tools with your existing workspace and project permissions.

Questions this page should answer

  1. How do I create a token safely?
  2. Which MCP endpoint should my client use?
  3. How do I confirm tools are available in Cursor, Claude, or another client?
  4. What should I do when a token or project selection fails?

Before you connect

  • In Atomic, open Profile settings and select MCP.
  • Create an MCP token and copy the raw atmcp_... value immediately. Atomic only shows it once.
  • Use the MCP endpoint:

What this page gives you

  • Your Atomic MCP endpoint.
  • Personal MCP token creation and revocation.
  • Token list with masked values and status.
  • The configuration pattern for clients that support remote HTTP MCP.

Important: Treat MCP tokens like passwords. Store them in your client config or secret manager, not in shared docs or chat threads.

How project access works

MCP tokens inherit your user permissions. They cannot access projects you cannot already open in Atomic. You do not need to put a project id in your client configuration. When a tool runs:
  1. Atomic uses the only accessible project automatically when your account has access to one project.
  2. If your account has access to multiple projects, Atomic returns a project chooser with each project id, name, domain, and workspace.
  3. Ask the AI client to use the right project and rerun the tool with projectId in the tool arguments.

Cursor

Cursor can connect directly to Atomic’s HTTP MCP endpoint. Open Cursor MCP settings and add this server to your mcp.json:
After saving, refresh the atomic-ai MCP server in Cursor. The tools list should load without any project-specific header.

Claude Desktop

Claude Desktop expects local stdio MCP servers in claude_desktop_config.json. Use mcp-remote to bridge Claude Desktop to Atomic’s HTTP MCP endpoint. Add this server to ~/Library/Application Support/Claude/claude_desktop_config.json on macOS:
Fully quit and reopen Claude Desktop after editing the file.

Claude Code

Claude Code can add Atomic as a remote HTTP MCP server from the terminal:
Then run this inside Claude Code:
Confirm that atomic-ai is connected and that tools are available.

Other MCP clients

Use Atomic as a remote HTTP MCP server when the client supports HTTP MCP:
If the client only supports local stdio servers, use mcp-remote with the Claude Desktop pattern.

Test the connection

You can test the token and endpoint with curl:
A healthy response includes a tools array.

Quick setup checklist

  1. Create a fresh MCP token.
  2. Add the endpoint and Authorization header to your client.
  3. Restart or refresh the MCP client.
  4. Confirm the Atomic tools list appears.
  5. Run a harmless read-only tool first.
  6. If asked to choose a project, pass the returned projectId on the next call.

Troubleshooting

The client shows no tools

Check that the MCP token is active, the endpoint ends with /api/mcp, and the authorization header starts with Bearer .

Claude Desktop says the config is invalid

Use the command and args format with mcp-remote. Claude Desktop does not accept the direct url and headers format used by Cursor.

The tool asks you to choose a project

Choose the project by name or domain. The AI client should call the tool again with the returned projectId in the tool arguments.

A token was shared by mistake

Revoke it from Profile settings > MCP immediately and create a new token.

Keep in mind

  • MCP tokens inherit your user permissions.
  • Tokens do not grant access to projects you cannot open in Atomic.
  • Project selection is explicit when your account can access multiple projects.
  • Revoke old tokens when you stop using a client.

Where to go next