Claudesidian

An Obsidian plugin that integrates Claude AI directly into your vault. Chat with Claude in a sidebar, run writing commands on your notes, and let Claude read and modify your vault through tool calling.

---

First Start

1. Install the plugin

Copy these three files into your vault's plugin folder:

<your-vault>/.obsidian/plugins/claudesidian/
  main.js
  manifest.json
  styles.css

Then in Obsidian: Settings → Community plugins → turn off Restricted mode → enable Claudesidian.

2. Add your API key

Open Settings → Claudesidian and paste your Anthropic API key into the API key field. Get a key at console.anthropic.com.

Click Test to verify the connection.

3. Choose a model

Two models are available:

Model	Speed	Cost
Claude Sonnet 4.6	Fast	$3 / $15 per 1M tokens
Claude Haiku 4.5	Fastest	$1 / $5 per 1M tokens

You can also switch models directly from the chat sidebar at any time.

4. Create your instructions file (optional but recommended)

Go to Settings → Claudesidian and click Create CLAUDE.md. This creates a CLAUDE.md file at your vault root and opens it for editing.

Fill in the template to tell Claude about your vault — its purpose, your writing style preferences, formatting rules, and any behaviours you want to enforce. Claude reads this file automatically on every request.

You can also place a CLAUDE.md inside any subfolder. Claude will merge instructions from the vault root through every parent folder of the active note, with more specific (local) instructions taking priority.

5. Set a spending limit (optional)

In Settings → Claudesidian, set a Monthly spending limit in dollars. Claude will stop accepting new messages once the limit is reached. The counter resets automatically on the first of each month. You can reset it manually at any time with the Reset usage button.

---

Features

Chat sidebar

Open the chat panel from the ribbon icon (💬) or via Command Palette → Open Claude Chat.

• Model switcher — change between Sonnet and Haiku without leaving the chat
• Attach note — click the paperclip button to attach the currently open note as context for your next message. The note name appears as a chip; click × to detach before sending
• Usage bar — shows current monthly spend vs your limit as a thin progress bar under the header. Turns red when the limit is reached
• Token count — each Claude response shows the number of output tokens at the bottom of the bubble
• Clear — trash icon clears the conversation history (session only, not persisted)
• Settings shortcut — gear icon opens the settings page directly

Messages support full Markdown rendering — headings, bold, code blocks, lists, and links all display correctly in Claude's responses.

Claude uses vault tools silently in the background. Notices appear when files are created or modified.

Writing commands

Three commands are available via the Command Palette (Cmd/Ctrl+P):

Command	What it does
Continue writing	Takes text before the cursor and streams a continuation
Summarize note	Summarizes the full content of the current note
Improve / rewrite selection	Rewrites the selected text while preserving its meaning

All three commands open a preview modal before applying any changes:

• The original text is shown on top
• Claude's suggestion streams in below in real time
• Accept — applies the change to the editor
• Retry — generates a new suggestion
• Cancel — discards and closes

Vault tools (agentic loop)

When asked, Claude can interact with your vault directly:

Tool	What it does
list_files	Lists files in a folder
read_note	Reads the full content of a note
create_note	Creates a new note with given content
update_note	Replaces the content of an existing note
search_notes	Full-text search across all notes
get_vault_structure	Returns the folder tree

CLAUDE.md instruction system

Claude loads instructions from CLAUDE.md files on every request:

1. CLAUDE.md at vault root (global instructions)
2. CLAUDE.md in each parent folder of the currently active note (local overrides)

Files are merged from global → local. Changes take effect immediately.

---

How it works

User message
    │
    ▼
Build system prompt
  ├─ Base instructions
  ├─ CLAUDE.md hierarchy (vault root → active note's parent folders)
  └─ Custom system prompt (from settings)
    │
    ▼
Anthropic API (streaming)
    │
    ├─ Text chunks → displayed incrementally in the chat bubble
    │
    └─ Tool calls (if any)
         ├─ Execute against app.vault
         ├─ Show Obsidian Notice
         └─ Feed result back → loop until Claude stops calling tools
    │
    ▼
Usage recorded (tokens → dollars, persisted monthly)

---

Settings reference

Setting	Description
API key	Your Anthropic API key. Stored in plugin data, never logged
Model	Claude Sonnet 4.6 or Haiku 4.5
Custom system prompt	Extra instructions appended to every request
Monthly spending limit	Block requests above this dollar amount (0 = no limit)
Current usage	Dollars spent this calendar month
Reset usage	Manually zero the usage counter
CLAUDE.md	Create or open the vault instruction file
Test connection	Send a ping to verify your API key

---

Privacy & security

• The API key is stored via Obsidian's plugin data (data.json) and is never logged or exposed
• Note content is only sent to Anthropic when you explicitly attach a note or Claude calls a vault tool
• data.json is listed in .gitignore so it won't be committed to version control
• The plugin is desktop-only (isDesktopOnly: true)