Obsidian Tools

An agentic Obsidian vault manager. Ask questions in natural language, search across notes semantically, explore wikilinks, transcribe meeting recordings, manipulate and organize files and metadata, all conversationally. Works as an Obsidian sidebar plugin, a CLI chat agent, or an HTTP API.

What It Does

Your vault gets indexed into a vector database. An LLM agent then uses MCP tools to search, read, and modify your notes — combining semantic understanding with keyword matching to find what you need.

Search & Discovery — Hybrid semantic + keyword search with BM25 scoring, Reciprocal Rank Fusion, HyDE (Hypothetical Document Embeddings) for question-type queries, cross-encoder reranking, and per-source diversity limits. Link graph traversal (backlinks, outlinks), frontmatter queries, date range filtering, folder browsing.

Vault Management — Read, create, move, merge files; edit specific markdown sections by heading; update frontmatter fields. Batch operations (create, move, merge, frontmatter update) with a built-in confirmation flow — the agent previews affected files and waits for approval before executing.

File Readers — Audio transcription via Whisper, image description via vision model, Office document extraction (.docx, .xlsx, .pptx), PDF text extraction, all auto-dispatched through read_file

AI-Powered Analysis — Summarize notes via LLM. Research pipeline: extract topics from a note or research an ad-hoc subject, gather findings from web search + vault search + page extraction, synthesize into a ## Research section or a new note.

Integrations — Web search via DuckDuckGo, interaction logging to daily notes, persistent user preferences

How It Works

1. Indexer scans your vault and creates embeddings in ChromaDB (using nomic-embed-text-v1.5), splitting notes by headings, paragraphs, and sentences with heading hierarchy prefixes ([Note > Section > Subsection]) and cross-section overlap for continuity
2. Search pipeline combines semantic search (ChromaDB) with BM25 keyword scoring via Reciprocal Rank Fusion. Question-type queries get HyDE augmentation (LLM generates a hypothetical answer, which is embedded alongside the original query). Results are reranked by a cross-encoder model and deduplicated per source.
3. MCP Server exposes 20 tools for searching, reading, and modifying vault content
4. LLM Agent (powered by Fireworks AI) orchestrates the tools to answer your questions
5. Interfaces — chat in Obsidian via the sidebar plugin, from the terminal via the CLI agent, or programmatically via the HTTP API

Requirements

• Python 3.11, 3.12, or 3.13 (not 3.14 — onnxruntime doesn't have wheels yet)
• Fireworks AI API key — required for the chat agent and audio transcription

Quick Start

git clone https://github.com/glibalien/obsidian-tools.git
cd obsidian-tools

# macOS / Linux
./install.sh

# Windows (PowerShell)
.\install.ps1

The installer will:

1. Find or help you install a compatible Python (resolves the real binary, not pyenv shims)
2. Create a virtual environment and install dependencies
3. Walk you through .env configuration (vault path, API key, etc.)
4. Optionally install background services (API server + vault indexer)
5. Optionally run the initial vault index

That's it — once installed, open Obsidian and start chatting, or run the CLI agent with python src/agent.py.

brew install pyenv

# Add pyenv to your shell (add these to ~/.zshrc or ~/.bashrc)
echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.zshrc
echo 'command -v pyenv >/dev/null || export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.zshrc
echo 'eval "$(pyenv init -)"' >> ~/.zshrc
source ~/.zshrc

pyenv install 3.12.8

git clone https://github.com/glibalien/obsidian-tools.git
cd obsidian-tools

# If using pyenv, create the venv with the real binary (not the shim):
$(pyenv which python3.12) -m venv .venv
# Otherwise:
python -m venv .venv

source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env

VAULT_PATH=~/Documents/your-vault-name
CHROMA_PATH=./.chroma_db
FIREWORKS_API_KEY=your-api-key-here
FIREWORKS_MODEL=accounts/fireworks/models/gpt-oss-120b
API_PORT=8000
INDEX_INTERVAL=60

Usage

1. Index your vault

Before searching, build the vector index:

python src/index_vault.py

The indexer is incremental — subsequent runs only process files modified since the last run and prune deleted files. Use --full for a complete reindex.

To keep the index up to date automatically, see Running as a Service.

2. Choose your interface

Obsidian Plugin (recommended for daily use)

The plugin/ directory contains a chat sidebar that connects to the API server. It includes a built-in batch operation confirmation system — when the agent wants to modify many files at once, the plugin renders a preview of affected files with Confirm/Cancel buttons so you stay in control.

cd plugin && npm install && npm run build

# Copy to your vault (adjust path)
mkdir -p ~/Documents/your-vault/.obsidian/plugins/vault-agent
cp manifest.json main.js styles.css ~/Documents/your-vault/.obsidian/plugins/vault-agent/

Enable "Vault Agent" in Obsidian Settings > Community Plugins. The API server must be running (see below).

CLI Agent (for terminal users)

python src/agent.py

HTTP API (for programmatic access)

python src/api_server.py

The server binds to 127.0.0.1:8000 (localhost only):

curl -X POST http://127.0.0.1:8000/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Summarize this note", "active_file": "Projects/Marketing.md"}'

Sessions are keyed by active_file — same file continues the conversation, different file starts a new one.

MCP Client (Claude Code, etc.)

Copy .mcp.json.example to .mcp.json and update the paths:

{
  "mcpServers": {
    "obsidian-tools": {
      "command": "/path/to/obsidian-tools/.venv/bin/python",
      "args": ["/path/to/obsidian-tools/src/mcp_server.py"]
    }
  }
}

Customizing for Your Vault

The agent's system prompt describes your vault's folder layout and frontmatter conventions. The default is tuned to the author's vault and almost certainly doesn't match yours. An uncustomized prompt wastes tokens on failed lookups and gives worse results.

The installer copies system_prompt.txt.example to system_prompt.txt (gitignored). Edit it:

1. Update the Vault Structure section with your actual folder layout and frontmatter conventions
2. Adjust the Interaction Logging section if you don't use daily notes
3. Keep the Choosing the Right Tool and Available Tools sections as-is — they apply universally

Available Tools

Running as a Service

The installer can set up background services automatically. If you prefer manual setup:

mkdir -p ~/.config/systemd/user

for f in services/systemd/*.service services/systemd/*.timer; do
    sed -e "s|__PROJECT_DIR__|$PWD|g" \
        -e "s|__VENV_PYTHON__|$PWD/.venv/bin/python|g" \
        -e "s|__INDEX_INTERVAL__|60|g" \
        "$f" > ~/.config/systemd/user/$(basename "$f")
done

systemctl --user daemon-reload
systemctl --user enable --now obsidian-tools-api
systemctl --user enable --now obsidian-tools-indexer-scheduler.timer

# Check status
systemctl --user status obsidian-tools-api

# View logs
journalctl --user -u obsidian-tools-api -f

# Restart
systemctl --user restart obsidian-tools-api

for f in services/launchd/*.plist; do
    sed -e "s|__VENV_PYTHON__|$PWD/.venv/bin/python|g" \
        -e "s|__PROJECT_DIR__|$PWD|g" \
        -e "s|__USERNAME__|$(whoami)|g" \
        -e "s|__INDEX_INTERVAL_SEC__|3600|g" \
        "$f" > ~/Library/LaunchAgents/$(basename "$f")
done

launchctl load ~/Library/LaunchAgents/com.obsidian-tools.api.plist
launchctl load ~/Library/LaunchAgents/com.obsidian-tools.indexer.plist

# Check status
launchctl list | grep obsidian-tools

# View logs
tail -f ~/Library/Logs/obsidian-tools-api.log

$xml = (Get-Content services\taskscheduler\obsidian-tools-api.xml -Raw) `
    -replace '__VENV_PYTHON__', "$PWD\.venv\Scripts\python.exe" `
    -replace '__PROJECT_DIR__', "$PWD"
Register-ScheduledTask -TaskName "ObsidianToolsAPI" -Xml $xml

$xml = (Get-Content services\taskscheduler\obsidian-tools-indexer.xml -Raw) `
    -replace '__VENV_PYTHON__', "$PWD\.venv\Scripts\python.exe" `
    -replace '__PROJECT_DIR__', "$PWD" `
    -replace '__INDEX_INTERVAL__', '60'
Register-ScheduledTask -TaskName "ObsidianToolsIndexer" -Xml $xml

# Check status
Get-ScheduledTask | Where-Object TaskName -like 'ObsidianTools*'

Uninstall

./uninstall.sh        # macOS / Linux
.\uninstall.ps1       # Windows

Your .env and .chroma_db/ are preserved.

src/
├── mcp_server.py        # FastMCP server — registers tools from submodules
├── api_server.py        # FastAPI HTTP wrapper with session management
├── agent.py             # CLI chat agent with tool result continuation
├── config.py            # Shared configuration
├── chunking.py          # Structure-aware markdown chunking (headings, paragraphs, sentences)
├── bm25_index.py        # In-memory BM25 keyword index (lazy singleton from ChromaDB docs)
├── hybrid_search.py     # Semantic + BM25 search with RRF, HyDE, reranking, source diversity
├── search_vault.py      # Search interface
├── index_vault.py       # Vault indexer (incremental, parallel chunking, batched upserts)
├── log_chat.py          # Daily note logging
├── services/
│   ├── chroma.py        # ChromaDB connection, embedding helpers, cross-encoder reranker
│   ├── compaction.py    # Tool message compaction for token management
│   └── vault.py         # Path resolution, response helpers, utilities
└── tools/
    ├── files.py         # File operations (read, create, move, merge, batch)
    ├── frontmatter.py   # Frontmatter queries and updates
    ├── links.py         # Backlinks, outlinks, folder comparison
    ├── preferences.py   # User preferences
    ├── search.py        # Vault search, web search
    ├── editing.py       # Section and position-based editing
    ├── utility.py       # Interaction logging
    ├── readers.py       # File type handlers (audio, image, Office docs, PDF)
    ├── summary.py       # LLM-powered file summarization
    └── research.py      # Agentic research pipeline (extract → search → synthesize)

plugin/                  # Obsidian chat sidebar (optional)
services/                # Service templates (systemd, launchd, Task Scheduler)

License

MIT