MCP client — connecting external tool servers

What this accomplishes

Connect a Model Context Protocol (MCP) tool server to Miton and use its tools from any lens.

Miton includes a built-in Model Context Protocol (MCP) client. Connect any MCP-compatible tool server and its tools become available to the agent in Chat, Writer, Canvas, Code, and Automation.

Current availability

This page documents behaviour verified in Miton development builds. Miton is still pre-release and is not publicly downloadable; invited beta users will receive a signed desktop build with automatic updates.

Before you start

• A Miton development build and an active project.
• A connected cloud or local model, plus any credentials or local runtime required by this task.

What is MCP?

MCP is an open protocol for exposing tools to AI agents. A tool server is any process that speaks the MCP protocol and exposes one or more tools. The agent calls tools by name; the server executes them and returns results.

Tool servers can expose anything: a database, a web scraper, an internal API, a file indexer, a custom search engine, a code runner, or a hardware sensor. If you can write a server that speaks MCP, Miton can use it.

Connecting a tool server

1. Open Settings → MCP → Add server
2. Enter the server’s transport URL. Miton supports:
   • stdio — for local processes (e.g. npx @modelcontextprotocol/server-filesystem)
   • sse — for Server-Sent Events (local or remote HTTP servers)
3. Click Connect. Miton performs tool discovery — it asks the server which tools it exposes and what their schemas are.
4. The discovered tools appear in Settings → MCP → [server name] → Tools. Toggle individual tools on or off per server.

Using MCP tools in a session

Once a server is connected, the agent can use its tools automatically. You do not need to mention the server by name. If the agent determines a tool is useful for the task, it calls it.

To force a specific tool, use the /tool skill or mention it explicitly: “Use the database-query tool to find all users created this month.”

Tool call transparency

Every MCP tool call appears in the transcript as a tool card — the tool name, the input arguments, and the result. In confirm mode (mode selector set to confirm), Miton pauses before every tool call and asks you to approve or reject it.

Permissions

MCP tool calls go through the permission system. By default:

• read tools (tools that only return data) are permitted in all modes
• write tools (tools that modify state) require autoedit or fullaccess mode, or explicit approval in confirm mode

You can override permissions per tool in Settings → MCP → [server name] → Tools → [tool name] → Permissions.

Example: file system server

npx @modelcontextprotocol/server-filesystem /path/to/directory

Connect it via stdio in Settings → MCP. The agent can now read and write files in the target directory using the MCP file system tools — useful for giving the agent access to directories outside your project root.

What success looks like

Connect a Model Context Protocol (MCP) tool server to Miton and use its tools from any lens. The result remains attached to the active project when you move to another workspace.

Common failures and recovery

• Confirm the correct project and model are active.
• Read the exact error and the requested permission before retrying.
• Retry only the failed step so completed project work remains intact.

Related next action

• Agent tool ergonomics — when to use MCP vs first-party tools vs agent-ergonomic CLIs (including AXI) via bash; MCP is not replaced by AXI (ADR-0218)
• Security & Permissions — how the permission system controls tool access
• Skills registry — for reusable task definitions that don’t need a full MCP server