MCP client configuration

Two protocol eras are served at once. A client that speaks 2026-07-28 needs no configuration and no handshake: it discovers the server with server/discover and states its version in every request. A handshake client uses 2025-11-25, and 2025-06-18 remains supported for existing clients.

Claude DesktopCopy Anchor Link

Edit claude_desktop_config.json — ~/Library/Application Support/Claude/ on macOS, %APPDATA%\Claude\ on Windows — and restart Claude Desktop:

{
  "mcpServers": {
    "coding-tools": {
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

CodexCopy Anchor Link

[mcp_servers.coding_tools]
command = "uvx"
args = ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]

Claude CodeCopy Anchor Link

{
  "mcpServers": {
    "coding-tools": {
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

Or from the command line:

claude mcp add coding-tools -- uvx coding-tools-mcp --stdio --workspace /path/to/repo

CursorCopy Anchor Link

{
  "mcpServers": {
    "coding-tools": {
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

VS CodeCopy Anchor Link

GitHub Copilot reads .vscode/mcp.json in the workspace:

{
  "servers": {
    "coding-tools": {
      "type": "stdio",
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

For the local HTTP server instead, use "type": "http" with "url" and "headers" keys.

WindsurfCopy Anchor Link

Edit ~/.codeium/windsurf/mcp_config.json and restart Windsurf:

{
  "mcpServers": {
    "coding-tools": {
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

Gemini CLICopy Anchor Link

One command, project scope by default:

gemini mcp add coding-tools -- uvx coding-tools-mcp --stdio --workspace /path/to/repo

Or edit ~/.gemini/settings.json (user scope) or .gemini/settings.json (project scope):

{
  "mcpServers": {
    "coding-tools": {
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

For the local HTTP server, transport http replaces the command entirely:

gemini mcp add --transport http coding-tools http://127.0.0.1:8765/mcp

Authenticated endpoints take --header "Authorization: Bearer <token>". Keep the server name free of underscores: Gemini CLI derives tool names from it and splits on the first one.

Continue, Cursor, Cline, And Generic HTTP ClientsCopy Anchor Link

Configure a Streamable HTTP MCP server at:

http://127.0.0.1:8765/mcp

The server is designed for local loopback use. Do not bind it to a public interface without external authentication and sandboxing.

Remote MCPCopy Anchor Link

For remote MCP clients, keep the server on loopback and expose it through an HTTPS tunnel with authentication. The fixed tool set includes mutation and command execution:

CODING_TOOLS_MCP_AUTH_MODE=bearer \
integrations/tunnels/tunnel.sh cloudflared /path/to/repo

Configure the remote MCP client with:

URL: https://<tunnel-host>/mcp

Static bearer-token auth is available for clients that support custom Authorization headers. OAuth-aware MCP clients can use --oauth-mode, which publishes protected-resource and authorization-server discovery plus RFC 7591 dynamic registration and a PKCE authorization flow. Clients that support neither require an external authenticated proxy. See Remote access.

ChatGPTCopy Anchor Link

ChatGPT is a cloud client: it cannot launch a local process, so it needs the authenticated HTTPS tunnel from Remote MCP, and its custom connectors offer OAuth or no authentication — there is no static bearer header to enter.

CODING_TOOLS_MCP_AUTH_MODE=oauth integrations/tunnels/tunnel.sh cloudflared /path/to/repo

In ChatGPT, enable developer mode (Settings → Connectors → Advanced settings), then create a custom connector pointing at https://<tunnel-host>/mcp. If ChatGPT discovers the OAuth endpoints itself, the authorization page asks for the password the script printed. If the connector form asks for details instead, fill them in:

• Authorization URL: https://<tunnel-host>/oauth/authorize
• Token URL: https://<tunnel-host>/oauth/token
• Client ID and secret: from a pre-registered client — start the tunnel with CODING_TOOLS_MCP_OAUTH_CLIENT_ID, CODING_TOOLS_MCP_OAUTH_CLIENT_SECRET, and CODING_TOOLS_MCP_OAUTH_REDIRECT_URIS set to the redirect URI ChatGPT shows.

GrokCopy Anchor Link

grok.com is also cloud-only. Start the same OAuth tunnel, then add a custom connector at grok.com/connectors (Settings → Connectors) with https://<tunnel-host>/mcp; Grok completes the OAuth flow in a popup.

The xAI API takes the static bearer tunnel instead — a request's remote MCP tool entry pins the server and the header:

{
  "server_url": "https://<tunnel-host>/mcp",
  "server_label": "coding_tools",
  "authorization": "<CODING_TOOLS_MCP_AUTH_TOKEN>"
}

Only Streamable HTTP and SSE transports are supported on Grok's side, which is what the tunnel already speaks.