Connect an AI assistant

loch mcp starts an MCP server (stdio transport). It uses the endpoints and API keys of your loch configuration.

Before you start: install loch, then run loch config, loch use and loch login. Use the full path of the binary (which loch): desktop apps often do not see your PATH.

Claude Code

claude mcp add --scope user loch /usr/local/bin/loch mcp

Check with claude mcp list.

Codex

codex mcp add loch -- /usr/local/bin/loch mcp

Or in ~/.codex/config.toml:

[mcp_servers.loch]
command = "/usr/local/bin/loch"
args = ["mcp"]

OpenCode

In opencode.json (project) or ~/.config/opencode/opencode.json (global):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "loch": {
      "type": "local",
      "command": ["/usr/local/bin/loch", "mcp"],
      "enabled": true
    }
  }
}

LM Studio

In LM Studio, open Program → Install → Edit mcp.json, then add:

{
  "mcpServers": {
    "loch": {
      "command": "/usr/local/bin/loch",
      "args": ["mcp"]
    }
  }
}

Choose a model that supports tool calls.

Try it

Ask your assistant: "List my loch endpoints, then the stacks on the first one."

Good to know

• The assistant can do only what your API key allows. Use a key with limited scopes for daily work.
• Passwords are never returned to the assistant. To reset one, run loch services reset-password yourself.
• Deleting a database and publishing ports are CLI-only. So are deploying functions, their secrets and credentials: an assistant only invokes the functions marked --mcp-invocable.
• loch ai:howto prints these instructions in the terminal.

Full list of tools: MCP tools.