Your notes are useful to you. They'd be more useful if the AI in your editor could read them too, so that "how did I deploy the worker last time?" gets answered from your runbook, not a generic guess.
That's what the Model Context Protocol (MCP) is for. It's a small, open standard for exposing tools and data to AI clients like Claude Code and Cursor, so they can call into your stuff in a structured way instead of you copy-pasting context every time.
The shape of it
MCP is a client/server protocol. Your tool runs a server that advertises a few capabilities; the AI client connects and uses them.
- Tools: functions the model can call (e.g.
search_notes(query)). - Resources: data the model can read (files, records, docs).
- Prompts: reusable prompt templates the server provides.
The client handles discovery: it asks the server what it offers, then the model decides when to call a tool, passing arguments and getting structured results back.
Claude Code ──connect──▶ your MCP server
◀─advertise─ tools: [ search_notes, get_note ]
──call──────▶ search_notes("staging db password")
◀─results─── [ { title, snippet, id }, ... ]Why read-only is the right default
It is tempting to expose write tools too: let the model file notes, edit secrets, run commands. Resist that by default. A read-only server has a small, boring blast radius: the worst case is the model reads something, not that it changes or deletes something on a hallucinated tool call.
Start read-only. Add write capabilities deliberately, one at a time, with confirmation, not as a convenience.
What it looks like for a notes CLI
inode ships a read-only MCP server. Point an MCP client at it and the model gets exactly two abilities: search the knowledge base by meaning, and fetch a specific entry. Nothing writes, nothing deletes, and sensitive values stay masked the same way they are in the terminal.
The result is that your editor's assistant can answer from your real notes, commands, and runbooks, while the knowledge base stays the single source of truth on your machine, not something a model can quietly rewrite.
MCP is the thin, well-defined seam between "the AI in my editor" and "my own tools." Keep the seam small and read-only, and it's almost all upside.