Claude Code is a coding agent, but its core skills — read files, search, make targeted edits, run commands — are exactly what you want over a folder of Markdown notes. The interesting part is connecting it so the agent has real tools instead of just cat and hope.

The plain-files approach

Because Markdown notes are ordinary files, Claude Code works on them out of the box: open it in your notes folder and it can search, read, and edit. Three habits make this safe rather than terrifying:

Keep the folder in Git. Then every agent edit is a diff you can read and revert, which is also just good practice for notes. Our Git sync guide covers the setup.
Tell it the house rules. An AGENTS.md at the root of the folder is read like CLAUDE.md in a repo: how notes are named, what frontmatter means, what it should never touch.
Read before writing. Ask it to summarize what it found before letting it edit, the same way you'd review a plan before a refactor.

The limitation: the agent sees files, not the app. It doesn't know your tasks, your Properties, or your task IDs — unless you give it a server that speaks for them.

The MCP approach, with the Octarine CLI

Octarine ships a CLI with a built-in MCP server, so the agent gets notes-aware tools instead of file operations:

claude mcp add octarine -- octarine mcp --workspace "Work Notes"

That gives Claude Code access to the same commands the CLI exposes — searching notes, reading and editing them, and working with tasks and their statuses. Two properties matter:

It runs locally. The server talks to the client that started it, over standard input and output; note content doesn't go to an Octarine service. What the agent does with results depends on Claude Code and the AI provider behind it, which is the same boundary as any Claude Code session.
The desktop app doesn't need to be open. The CLI is a separate program with its own access to the workspace.

There's also octarine agent-guide, which prints a briefing the agent can read — how to find notes, read before editing, and recover from revision conflicts — useful pasted into AGENTS.md. And help --json describes every command and error code, so the agent doesn't have to guess.

What this is good for

Bulk reorganization — moving notes into folders, normalizing titles and frontmatter, fixing link syntax across hundreds of files.
Working a task list — agents can read a task, act, move it through statuses, and comment as they ship; stable task IDs are safe in commit messages.
Summarizing and synthesizing — "read everything tagged insomnia-study and draft a summary note" is a retrieval job agents do well.
Meeting and project notes — turning a messy capture into structured notes with tasks, on your machine.

What to watch for

An agent with edit rights is an agent that can be wrong at scale. The mitigations are boring and effective: Git history, a written brief, small first tasks, and a diff review before you push. Treat the agent like a well-meaning collaborator who never gets tired but also never double-checks its own certainty.

Sources and review notes

The Octarine CLI, its MCP server, agent-guide, and help --json are documented on the CLI page and in Octarine's docs.
Claude Code behavior (CLAUDE.md, MCP support) is described in Anthropic's documentation.
The privacy boundary — the MCP server talks only to its client — is stated on the CLI page; verify against current docs before relying on it.