Connecting MCP servers

Bring your own tools: connect any Model Context Protocol server and the AI can use it right from Kerlig.

MCP (Model Context Protocol) is an open standard for giving AI models access to tools. Thousands of servers exist for apps like Notion, Linear, Grafana, databases, and more. Add servers in Settings → MCP.

Local servers

A local server is a program Kerlig starts on your Mac. Add one with the + button, choose Local, and fill in:

• Command: the program to run, e.g. npx. Kerlig resolves commands using your shell's PATH, so anything that works in your terminal works here.
• Arguments: space- or line-separated. Quote arguments that contain spaces. For example: -y @modelcontextprotocol/server-everything
• Environment variables: API keys and other secrets the server needs. Values are stored in Kerlig's encrypted database and shown masked.

Before a local server first runs, Kerlig shows a trust prompt with the exact command. Local servers are programs, so only add ones you trust. The prompt reappears if the command changes.

Remote servers

A remote server is a URL. Choose Remote, paste the server's address, and pick how to authorize:

• None: for open servers.
• Headers: static credentials, e.g. an API token header.
• OAuth: click Connect and sign in with your browser. Kerlig handles the rest and refreshes tokens automatically.

Import from JSON

Copy a server configuration in the common mcpServers JSON format (most MCP documentation provides one) and use Import from clipboard. Kerlig creates the servers for you.

Managing servers

Each server shows a status dot: running, starting, needs sign-in, or failed, with the server's own error output shown verbatim when something goes wrong. Placeholder-named servers adopt the name the server reports about itself, and you can rename them anytime. You can disable individual tools per server to reduce token usage, and Kerlig offers at most 60 MCP tools to the model per request.