Tutorial: Pair with an Agent

Get an AI agent working inside your notes, three ways: a conversation in a note's terminal, one-shot prompt fences, and capture from anywhere.

This builds on Agents and Ledge. The examples use Claude Code, but any MCP-speaking CLI works.

1. Connect the agent

Install the shell command with "Install Shell Command (ledge)" in the palette, then connect the agent to Ledge's MCP server from a terminal:

claude mcp add ledge -- ledge mcp

2. Talk to an agent in a note

Open a note you are working on, press ⌃for its terminal drawer, and startclaude`.

The terminal belongs to the note, so the agent already knows what "here" means. Ask it to "read this note and tell me what's missing", "add a section comparing the two options above", or "find my other notes that mention this design". No paths, and no copying text into a chat window. Edits land in the file you are looking at, live.

Use this for working sessions: you write, the agent reads and reacts, and everything it adds is in your note when the session ends rather than in a chat log.

3. Turn repeated asks into prompt fences

For an instruction you give repeatedly, write it into the note as a prompt fence and it becomes a button. A meeting note might end with:

Extract every action item from this note into a checklist under a new "Actions" heading, with owners in bold.

Give that fence the prompt language and after every meeting it is one ⌘↩.

The fence runs one-shot: the agent acts and reports rather than asking questions, so write the instruction with the decisions already made.

Fences work in templates too (Daily Notes and Templates). A template carrying a fence gives every stamped note a built-in agent action, like the morning briefing in Tutorial: A Daily Workflow.

4. Capture from anywhere

An agent that can run shell commands can work your notes with the CLI alone, with no MCP setup (The ledge CLI).

Tell a Claude Code session in any project to "append what we just decided to my Decisions note" and ledge append gets it there. ledge search and ledge cat let it check your notes before answering. Your notes become memory that outlives any one session.

The boundaries

Agents can read and write notes but never delete them. The trash and Undo stay yours, in the app.

A locked note's body is invisible to every agent surface (Note Locking). Lock the notes you want kept out of the loop.