Note Watcher

A tool that detects @ mentions in Obsidian markdown notes stored in Git and dispatches instructions to configured agents — like Claude Code — that can read, modify, and reorganize your notes directly.

Write @agent_name do something in any note, and Note Watcher dispatches the instruction to the named agent. The agent can edit files, create new notes, restructure content, or make any other changes to your vault. The original instruction is then replaced with a completion marker (an HTML comment, invisible in rendered markdown) so it is never reprocessed:

<!-- @done agent_name: do something
Agent response summary goes here.
/@done -->

The real work happens in the commit: the agent's changes to your vault are committed back to Git. The completion comment is just a record that the instruction was processed.

Modes of Operation

Requirements

• Python 3.10+

Installation

pip install notes-watcher

For development:

pip install -e ".[dev]"

Configuration

Copy the example config and edit it:

mkdir -p ~/.config/note-watcher
cp config.example.yml ~/.config/note-watcher/config.yml

The default config location is ~/.config/note-watcher/config.yml. You can override it with --config:

note-watcher watch --config /path/to/config.yml

Config reference

# Path to your Obsidian vault
vault: ~/Obsidian/MyVault

# Seconds to wait before processing after a file change
debounce_seconds: 1.0

# File patterns to ignore (glob syntax)
ignore_patterns:
  - "*.excalidraw.md"
  - ".trash/**"

# Agent definitions
agents:
  summarizer:
    type: echo        # Returns instruction text unchanged
  uppercase:
    type: uppercase   # Returns instruction text in uppercase
  word_count:
    type: command
    command: "wc -w"  # Runs a shell command, passes instruction via stdin

Agent types

System prompts

Command agents include a default system prompt that tells the agent about the vault, the note being processed, and how to respond. The default prompt is:

You are working in an Obsidian vault at {vault_path}. The user has left an instruction in the note at {file_path}. Read the note, then modify it as requested by the instruction. If the user asks for changes to a note without specifying which one, apply the changes to the same note that contains the instruction. After making your changes, commit them to git. Respond with a brief summary of what you did.

This works out of the box — you only need to configure a system prompt if you want different behavior.

Overriding the default prompt

To replace the default, set system_prompt inline or load it from a file with system_prompt_file. Setting either one completely replaces the default prompt.

agents:
  claude:
    type: command
    command: "claude --print"
    # Inline — replaces the default prompt entirely
    system_prompt: |
      You are a note-taking assistant working in {vault_path}.
      Edit the note at {file_path} as instructed.
      Do not commit changes — the caller will handle that.

  claude-from-file:
    type: command
    command: "claude --print"
    # Load from a file (path relative to the config file's directory)
    # Also replaces the default prompt entirely
    system_prompt_file: prompts/claude.md

You cannot set both system_prompt and system_prompt_file on the same agent — Note Watcher will raise an error if you do.

Template variables

System prompts (including the default) support template variables that are interpolated at dispatch time:

Environment variables

The resolved prompt is always passed to the command via the NOTE_WATCHER_SYSTEM_PROMPT environment variable. The following environment variables are always set for command agents:

Example: Using Claude Code as an agent

Configure a command agent that dispatches instructions to Claude Code. The default system prompt works well for Claude Code, so you only need to set the command:

agents:
  claude:
    type: command
    command: "claude --print --system-prompt \"$NOTE_WATCHER_SYSTEM_PROMPT\""

Claude Code runs with full access to your vault, so it can edit notes, create new files, and reorganize content — not just respond in a comment. Write @claude instructions in your notes:

@claude Summarize the key points of this meeting and add action items to my Tasks note

Daemon Mode

Daemon mode continuously watches your Obsidian vault for changes and processes @ mentions in real time.

Running manually

# Watch the vault specified in your config
note-watcher watch

# Override the vault path
note-watcher watch --vault ~/Obsidian/MyVault

# Enable verbose logging
note-watcher -v watch --vault ~/Obsidian/MyVault

Stop the daemon with Ctrl+C (SIGINT) or SIGTERM.

Installing as a macOS LaunchAgent

The included install script sets up Note Watcher as a LaunchAgent that starts on login and restarts on crash:

./scripts/install.sh

The script is idempotent and safe to run multiple times. It will:

1. Detect the note-watcher executable on your system
2. Generate a LaunchAgent plist from the included template
3. Install it to ~/Library/LaunchAgents/
4. Start the daemon

Logs are written to ~/Library/Logs/note-watcher/.

Uninstalling the LaunchAgent

./scripts/uninstall.sh

To also remove the log directory:

./scripts/uninstall.sh --clean

GitHub Action Mode

GitHub Action mode processes all pending @ instructions across the entire vault in a single batch run. This is useful for vaults stored in a Git repository.

CLI usage

note-watcher process --all --vault /path/to/vault

Setting up the GitHub Actions workflow

See examples/github-action/ for a complete, ready-to-copy example that uses Claude Code as the AI agent.

To set it up:

1. Copy examples/github-action/.github/ into your notes repository
2. Add a config.yml to your notes repo (see examples/github-action/config.example.yml)
3. Add your ANTHROPIC_API_KEY as a repository secret

The example workflow already includes permissions: contents: write, which is required for the action to push processed results back to your repository. If you write your own workflow, make sure to include this permission block.

The example workflow also includes a concurrency group keyed on the branch, with cancel-in-progress: false. This queues runs triggered by rapid, overlapping pushes instead of letting them race to commit at the same time — without it, a push that lands while a prior run is still processing can make that later run's commit fail. If you write your own workflow, include this too:

concurrency:
  group: note-watcher-${{ github.ref }}
  cancel-in-progress: false

The workflow triggers on any push that modifies .md files, processes all unprocessed @ instructions, and commits the agent's changes back to your repository. It uses [skip ci] to prevent infinite loops.

Upgrading an existing installation to fix overlapping-push failures

If you installed this workflow before the concurrency block above was added, a push with an @mention that lands while an earlier @mention push is still processing can make the later run fail (issue #28). To fix an existing installation:

1. Open the workflow file you copied into your notes repository (e.g. .github/workflows/note-watcher.yml).
2. Add the concurrency block shown above, at the top level of the workflow (alongside on: and permissions:, not inside jobs:):

   concurrency:
     group: note-watcher-${{ github.ref }}
     cancel-in-progress: false
   

   This part only lives in your own workflow file — the action can't set it on your behalf — so it always requires this manual edit, no matter how you reference the action below.
3. Check how your workflow references the action:
   • uses: britt/obsidian-notes-watcher@main — no further action needed. The other half of the fix (resyncing before processing) lives in action.yml itself and takes effect automatically on your next run once it's merged to main.
   • uses: britt/obsidian-notes-watcher@v0.4.x (or any other pinned tag/SHA) — bump the pin to a release that includes this fix (v0.4.5 or later) once one is published.
4. Commit and push the workflow change. No config or vault changes are needed.

See the Claude Code GitHub Actions documentation for more on setting up Claude Code in CI.

Arcade MCP Integration

Note Watcher can use Arcade as an MCP gateway to give Claude Code access to external services like Gmail, Google Drive, Google Calendar, Google Sheets, Google Docs, GitHub, and Slack. This lets your @claude instructions interact with services beyond the vault — for example, checking email, creating calendar events, or posting to Slack.

Arcade handles OAuth on your behalf: you authorize each service once from your local machine, and Arcade caches the tokens server-side. After that, the GitHub Action runner can use the same tokens without any interactive auth flow.

Setting up an Arcade MCP gateway

1. Create an Arcade account and note your API key
2. Go to the MCP Gateways dashboard and click Create MCP Gateway
3. Configure the gateway:
   • Name: A descriptive name (e.g. "Obsidian Note Watcher")
   • Slug: This becomes your gateway URL path (e.g. obsidian-notes-watcher gives you https://api.arcade.dev/mcp/obsidian-notes-watcher)
   • Authentication Mode: Use Arcade Headers for production — this requires an Authorization header with your API key and an Arcade-User-ID header with the user's email
4. Add the tools your agents need from the Arcade tool catalog. The following toolkits are supported:

You don't need to add every tool — only the ones your agents will use. You can always add more later from the dashboard.

Pre-authorizing OAuth tokens

Arcade uses a just-in-time OAuth flow: the first time a tool needs access to a service, the user must visit a URL in a browser to grant consent. Since GitHub Action runners are headless, you need to complete this authorization once from your local machine before the action can use those services.

The included scripts/authorize_arcade.py script handles this. It walks through each service, triggers the OAuth flow, opens your browser for consent, and waits for completion. Arcade then caches the tokens server-side, keyed by your email address.

Prerequisites

pip install arcadepy
export ARCADE_API_KEY=your-api-key

Authorize all services

python scripts/authorize_arcade.py you@example.com

The email address is your Arcade account email. It's used as the user_id for token storage — the same identity must be used when the MCP gateway is called from CI.

The script will open your browser for each service that hasn't been authorized yet. Grant consent for each one, and the script will detect completion automatically.

Authorize specific services only

python scripts/authorize_arcade.py you@example.com --services gmail slack google-calendar

Available services: github, gmail, google-drive, google-calendar, google-sheets, google-docs, slack.

Without auto-opening the browser

python scripts/authorize_arcade.py you@example.com --no-browser

This prints the authorization URLs instead of opening them, useful for remote machines or SSH sessions.

Re-authorization

You'll need to re-run the script if:

• An OAuth token is revoked or expires
• You add new services to your gateway
• Your Google Cloud OAuth app is in "testing" mode (tokens expire after 7 days — publish the app or set it to "internal" for Google Workspace orgs to avoid this)

Configuring the MCP server for Claude Code

Once your gateway is set up and tokens are cached, configure Claude Code to use it. In the GitHub Action workflow, add the MCP server before running Note Watcher:

- name: Configure Arcade MCP gateway
  run: |
    claude mcp add --transport http arcade \
      https://api.arcade.dev/mcp/your-gateway-slug \
      --header "Authorization: Bearer ${{ secrets.ARCADE_API_KEY }}" \
      --header "Arcade-User-ID: ${{ secrets.ARCADE_USER_ID }}"

Add these secrets to your repository:

Claude Code will then have access to all the tools in your gateway when processing @claude instructions. For example:

@claude Check my Gmail for any emails from the team about the Q4 report and summarize them here

@claude Create a Google Calendar event for our standup tomorrow at 10am

@claude Post a message in the #notes channel on Slack saying the weekly review is ready

Installing Skills for Claude Code

Skills are markdown files that teach Claude Code reusable behaviors — how to format notes, apply templates, follow your conventions, etc. Skills are checked into your notes repository, so they work automatically in both daemon mode and the GitHub Action.

Writing your own skills

Create a SKILL.md file inside a named directory under .claude/skills/:

your-notes-repo/
├── .claude/
│   └── skills/
│       ├── organize-notes/
│       │   └── SKILL.md
│       └── daily-journal/
│           ├── SKILL.md
│           └── templates/
│               └── journal-entry.md
├── Notes/
│   └── ...
└── config.yml

Each SKILL.md has YAML frontmatter and markdown instructions:

---
name: organize-notes
description: Organize notes into folders by topic and add backlinks
---

When asked to organize notes:
1. Group related notes into topic folders
2. Add `[[backlinks]]` between related notes
3. Update any broken links after moving files

Claude Code automatically discovers skills in .claude/skills/ and uses them when relevant. You can also reference one explicitly:

@claude Use the daily-journal skill to format today's standup notes

Installing plugins from a marketplace

You can install community-contributed plugins from a plugin marketplace instead of writing skills by hand. Run these commands locally from your notes repository:

# Install from the official Anthropic marketplace into the project
claude plugin install plugin-name@claude-plugins-official --scope project

# Commit and push so the GitHub Action picks it up
git add .claude/
git commit -m "Add plugin-name plugin"
git push

The --scope project flag writes the plugin configuration into the repo's .claude/ directory. Once committed and pushed, the plugin is available when the GitHub Action runs — no extra workflow steps needed.

To install from a third-party marketplace, add it first:

claude plugin marketplace add https://github.com/someone/their-marketplace
claude plugin install note-helper@their-marketplace --scope project

Then commit and push .claude/ as above.

Running Tests

pytest

With coverage:

pytest --cov=note_watcher

License

MIT