An AI toolbag for tools

A SQLite brain for your markdown vault

Index any Obsidian / Logseq / Foam vault into SQLite and expose it through an MCP server. Agents query with SQL & full-text search instead of reading raw files — ~8× fewer tokens, structured analytics, a self-improving optimizer.

Get started → See the configurator
37 vault_* MCP tools ~8× fewer tokens 100% local · no API key Obsidian · Logseq · Foam Adapts to your workflow
The problem

Reading raw files is the expensive path

An agent that reads 8 files to answer one question is 8× more expensive than one that runs a single query. vault-db-template compiles your notes into a queryable index so agents retrieve exactly what they need — structured retrieval over full-text search over raw reads, in that order.

The vault is the hub

Planning, code repos, infrastructure, remote services — everything radiates from the vault. Agents reach out from there; they don't pull you into their tools.

Notes become a compiled binary

Raw notes are immutable source. The DB is the compiled layer agents work against — Karpathy's LLM-wiki pattern. Rebuild any time; nothing is lost.

Local-first, agent-agnostic

SQLite on your machine — no cloud, no subscription. MCP means Claude Code, OpenCode, Cursor, and Antigravity all query the same vault.

How it works

Two processes, one SQLite file

The indexer (writer) and the MCP server (reader) never call each other — they share vault-index.db. Parse → compile → serve.

flowchart LR MD["markdown notes
(source)"] IDX["vault-indexer.py
(compiler)"] DB[("vault-index.db
(compiled)")] MCP["vault-mcp-server.py
37 MCP tools"] AG["agents"] MD -->|mtime scan| IDX IDX --> DB DB --> MCP MCP -->|fewer tokens, better answers| AG AG -.->|wiki synthesis| MD
From zero to querying

Three commands

bash
# 1 — install skills into every detected agent
$ ./bootstrap.sh

# 2 — pick features + write vdb-config.yaml (the TUI configurator)
$ vdb tui

# 3 — build the index, then query over MCP
$ vdb index --full --no-embed

Full walkthrough on the install page. Prefer a guided interview? Run /vdb-setup in your agent.

What's inside

Batteries, not lock-in

Adapts to your workflow. Every table group is opt-in — start with infrastructure, switch on roadmap, QA, and delegation when you need them, or add your own domain tables when the built-ins don't fit. Core retrieval is always on; the rest follows what you actually do.

Core retrieval

6 tools: search, chunk fetch, graph, status. FTS5 over heading-path chunks.

Infra / DevOps

10 domain tables: hosts, services, incidents, runbooks, certs, change log, cloud, repos.

Roadmap

Opt-in goals → epics → stories with write tools for agent dispatch.

QA impact graph

Opt-in service call graph + scenarios → blast-radius and coverage-gap analysis.

Delegation

Opt-in: hand a task to an agent in a git worktree (tmux or systemd), tracked in vault_agent_sessions.

Browse the tool catalog →