DuckVault-MCP

DuckVault-MCP v0.3.0 is a local RAG server for Obsidian Vaults and Markdown knowledge bases. It combines DuckDB vector search, local GraphRAG, Open Knowledge Format (OKF) support, and an offline interactive graph viewer behind an MCP interface.

• Quick Start

• Features

• CLI Reference

• MCP Tools

• OKF Authoring

• AI Agent Setup

• Troubleshooting

Quick Start — v0.3.0

1. Prerequisites

• Python 3.11 or newer

• uv

• An Obsidian Vault or directory containing Markdown files

• An absolute path to that directory

2. Install

From a local checkout:

git clone https://github.com/caron14/mcp-duckvault.git
cd mcp-duckvault
uv tool install .

Alternatively, install the packaged release:

uv tool install mcp-duckvault

3. Run the first sync

Use a dedicated database file for each Vault:

duckvault /absolute/path/to/vault \
  --db-path ~/.duckvault/my-vault.db \
  --sync-only

The first sync may download intfloat/multilingual-e5-small. Later starts load the model from ~/.duckvault/models and only re-index changed Markdown files.

4. Connect Claude Code

claude mcp add duckvault -- \
  duckvault /absolute/path/to/vault \
  --db-path ~/.duckvault/my-vault.db

claude mcp list

The registered command performs an incremental sync before starting the MCP server. Running the first sync separately avoids a model download during the initial MCP connection.

5. Try the retrieval tools

Ask your AI agent:

• “Find notes about customer revenue.”

• “Find notes tagged sales that discuss orders.”

• “Use find_related_notes for tables/orders.md.”

• “Use hybrid_search_notes to find customer metrics and linked notes.”

• “List OKF concepts with type table and tag sales.”

• “Explain the OKF concept tables/orders.”

6. Generate the offline graph viewer

duckvault /absolute/path/to/vault \
  --db-path ~/.duckvault/my-vault.db \
  --visualize \
  --output ./duckvault-graph.html

This writes:

duckvault-graph.html
duckvault-graph.json

Open the HTML file in a modern browser. It contains the graph data and Cytoscape.js, requires no backend or network connection, and does not embed the Markdown bodies.

Related MCP server: EzRAG MCP Server

Features

All indexed content and graph data remain local. Initial setup may download the DuckDB VSS extension and embedding model. The generated graph viewer makes no network requests; external resource links are opened only when selected.

How It Works

1. DuckVault scans Markdown files, excluding configured paths.

2. It parses YAML frontmatter, splits content at H1-H3 headings, and generates local embeddings.

3. It extracts document, link, heading, tag, folder, OKF type, resource, and citation graph records into DuckDB.

4. MCP tools query the vector index, traverse the graph, or combine both.

5. Visualization mode exports the persisted graph without re-parsing Markdown.

The existing search_notes vector behavior remains available alongside the GraphRAG and OKF tools.

CLI Reference

duckvault [OPTIONS] VAULT_PATH

VAULT_PATH must be an existing directory. Absolute paths are strongly recommended, and are required in AI agent configuration.

Common commands

Start the MCP server and filesystem watcher:

duckvault /absolute/path/to/vault

Use a custom database:

duckvault /absolute/path/to/vault --db-path ./vault-index.db

Choose both visualization output paths:

duckvault /absolute/path/to/vault \
  --visualize \
  --output ./artifacts/graph.html \
  --json-output ./artifacts/graph-data.json

MCP Tools

Hybrid results use:

final_score = 0.7 * vector_similarity + 0.3 * graph_score
graph_score = max(seed_vector_similarity / depth)

Document-oriented retrieval results include obsidian://open links.

Graph Visualization

The viewer initially displays documents and document-to-document relationships. The sidebar can enable:

• folders

• headings

• tags

• OKF types

• resources and citations

• dangling link targets

It also provides:

• title, path, concept ID, and tag search

• type, tag, directory, relation, and structural-node filters

• force-directed, concentric, breadth-first, circle, and grid layouts

• structured metadata and inbound/outbound relationships

• Obsidian and validated HTTP/HTTPS resource links

• orphan, dangling-link, duplicate-title, and high-degree diagnostics

The HTML enforces a Content Security Policy that prevents external connections. The JSON sidecar uses schema version 1.0.

OKF Authoring

Any Markdown file with leading YAML frontmatter containing type is indexed as an OKF Concept Document.

---
type: table
title: Orders
description: One row per customer order
resource: https://console.example.com/warehouse/orders
tags: [sales, finance]
timestamp: 2026-07-01T00:00:00Z
---

# Orders

Each order belongs to a [customer](./customers.md).
Revenue is defined by [weekly revenue](/metrics/weekly_revenue.md).

The Concept ID is the Vault-relative path without .md:

tables/orders.md -> tables/orders

Link resolution:

index.md and log.md are stored as navigation/history nodes, not OKF concepts. Frontmatter fields beyond the standard OKF fields remain available in node metadata.

Graph mapping

AI Agent Setup

Use the same absolute Vault path and database path in every configuration.

Claude Code

claude mcp add duckvault -- \
  duckvault /absolute/path/to/vault \
  --db-path /absolute/path/to/duckvault.db

Claude Desktop

Add this entry to the mcpServers object in ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "duckvault": {
      "command": "duckvault",
      "args": [
        "/absolute/path/to/vault",
        "--db-path",
        "/absolute/path/to/duckvault.db"
      ]
    }
  }
}

Gemini CLI

gemini mcp add --scope user duckvault \
  duckvault /absolute/path/to/vault \
  --db-path /absolute/path/to/duckvault.db

Verify with:

gemini mcp list

GitHub Copilot CLI

Start Copilot, run /mcp add, and register:

name: duckvault
command: duckvault
args: /absolute/path/to/vault --db-path /absolute/path/to/duckvault.db

Configuration

Database and model cache

Use a separate --db-path for each Vault. Synchronizing an unrelated Vault against the same database replaces the indexed document set.

Excluding files

Create .vaultignore at the Vault root. Patterns use glob matching. .obsidian and .trash are always excluded.

# Directories
private/
drafts/

# File patterns
archive-*.md

Troubleshooting

The first start is slow

The embedding model is loaded lazily and tried from the local cache first. If it is not cached, DuckVault downloads it from Hugging Face. Run --sync-only before registering MCP to complete this step separately.

MCP fails to connect

• Confirm duckvault is available in the same environment as the AI agent.

• Use absolute Vault and database paths.

• Run the configured command manually with --sync-only.

• Add --verbose; logs are written to stderr so MCP stdout remains valid.

Notes or graph records look stale

DuckVault normally re-indexes files when their MD5 changes. If parsing, schema, or embedding configuration changed between versions, remove the affected database and run a full sync again:

rm ~/.duckvault/my-vault.db
duckvault /absolute/path/to/vault \
  --db-path ~/.duckvault/my-vault.db \
  --sync-only

Back up the database first if it is needed for diagnostics.

The graph viewer is empty

• Confirm the sync completed successfully.

• Check that Markdown files are not excluded by .vaultignore.

• Verify that HTML and JSON output paths differ.

• Inspect the JSON sidecar to confirm stats.documents is greater than zero.

Broken links appear in diagnostics

This is expected for unresolved local links. DuckVault preserves them as link_target nodes so they can be found and repaired.

Development

uv sync --all-extras
uv run pytest
uv run black --check src tests
uv run isort --check-only src tests

The functional source of truth is SPECIFICATION.md.

Current Scope

v0.3.0 does not include LLM entity extraction, community detection, community summaries, Global GraphRAG, Neo4j export, GraphML/CSV export, or CI-enforced graph quality gates.

License

DuckVault-MCP is released under the MIT License. The bundled Cytoscape.js library is also MIT-licensed; its license text is included at src/mcp_duckvault/assets/CYTOSCAPE_LICENSE.txt.