Open-source local AI coding tools · Your files stay on your machine · Explore on GitHub ↗
SFI Agentby Shah Fahad Inc.
Read docsInstall SFI Agent ↗
Integrations

Your assistant, connected
to your local tools.

Use the guide that matches your client. Features and setup steps vary by app version and account plan.

ChatGPT custom MCP

Use this path only if your ChatGPT workspace/account supports custom MCP apps or connectors. Availability and write/action support can vary by plan and product surface.

  1. Start SFI Agent locally and configure the intended remote MCP endpoint.
  2. Enable a Cloudflare Tunnel or ngrok only if remote access is required.
  3. In the ChatGPT interface that supports custom MCP, open the connector/app configuration and add the public HTTPS MCP endpoint supplied by your host.
  4. Choose the authentication method configured by your server. Do not select “None” for privileged filesystem or terminal tools unless the endpoint is protected by an appropriate separate access policy.
  5. Save and test a harmless diagnostic tool before allowing edits or command execution.
Example endpoint shape only: https://your-host.example/mcp. Use the actual MCP endpoint reported by your installation; a website URL alone does not prove MCP compatibility.

Claude Desktop

Claude Desktop supports MCP workflows, but local stdio configuration and remote HTTP MCP configuration are different patterns. Use the one supported by your current Claude Desktop version and the SFI Agent release.

  1. For a local stdio setup, use the exact command and path printed by your installation or documented in the repository.
  2. For a remote setup, configure the server URL and authentication through the supported Claude interface, if available for your version.
  3. Restart or reload the client if required, then verify the server and tools appear.

Cursor and other MCP clients

Find the client's MCP server settings and select a transport compatible with the running SFI Agent server. Some clients expect a local command; others can connect to remote HTTP endpoints. Copy the correct configuration from your installed release rather than assuming all clients share the same format.

Verify the connection

  1. Check the local host status.
  2. Run a read-only diagnostic tool.
  3. List available tools and inspect their descriptions.
  4. Try reading a non-sensitive file inside the authorized workspace.
  5. Only then test edits, commands, and Git workflows.

If initialization or tool discovery fails, see troubleshooting and check HTTP status codes, redirects, content type, transport negotiation, and server logs.