holly-db

A portable, searchable knowledge graph for developers and their AI agents. Backed by SQLite with full-text and semantic search. Works with Claude Code, Cursor, Codex, or no agent at all — your knowledge stays with you.

Install

From GitHub Releases (recommended)

Download the latest binary for your platform from the Releases page:

Platform	File
Linux x86_64	holly-vX.Y.Z-x86_64-unknown-linux-gnu.tar.gz
macOS Intel	holly-vX.Y.Z-x86_64-apple-darwin.tar.gz
macOS Apple Silicon	holly-vX.Y.Z-aarch64-apple-darwin.tar.gz

Extract and place the holly binary somewhere on your PATH:

tar xzf holly-vX.Y.Z-*.tar.gz
mv holly ~/.local/bin/   # or /usr/local/bin/

Each release includes .sha256 checksums and Sigstore build attestations for verification.

From source

cargo install --path holly-cli

Quick Start

holly init                                # initialize database
holly remember "SQLite: zero ops, good enough for 100k nodes"
holly search "SQLite"
holly list --type decision
holly mcp-server                          # start MCP server for AI agents

Commands

Command	Alias	Description
holly init		Initialize database and download embedding model
holly remember "<text>"	mem	Quick capture (casual mode)
holly record --type <type> "<title>"	r	Structured record
holly search "<query>"	s	Hybrid FTS + semantic search
holly list	l	List nodes
holly get <id>	g	Get node details
holly edit <id>	update	Update a node
holly delete <id>	rm	Delete a node
holly connect <from> <to>		Connect two nodes
holly context		Export context for AI agents
holly audit		Health check
holly stats		Statistics
holly reindex		Backfill missing vector embeddings
holly event record <type>		Record lifecycle event
holly event list		List events
holly task create "<title>"		Create task
holly task start <id>		Start task
holly task complete <id>		Complete task
holly run start --task <id>		Start a run
holly run complete <id>		Complete a run
holly import --from <path>		Import from legacy Holly
holly mcp-server		Start MCP server (stdio transport)

MCP Integration

holly-db exposes all 21 tools as an MCP server over stdio, compatible with Claude Code, Cursor, and any MCP client.

Add to your .mcp.json:

{
  "mcpServers": {
    "holly-kg": {
      "command": "holly",
      "args": ["mcp-server"],
      "env": {
        "HOLLY_DB_PATH": "/path/to/your/.holly-db/holly.db",
        "HOLLY_AGENT": "claude-code",
        "HOLLY_LLM": "claude-sonnet-4-6"
      }
    }
  }
}

Database Location

Resolution order:

1. --db <path> flag
2. HOLLY_DB_PATH env var
3. Walk up directories for .holly-db/holly.db
4. ~/.holly-db/holly.db (global fallback)

Provenance

holly-db auto-stamps all writes with agent/user/LLM identity:

Env Var	Description
HOLLY_AGENT	Agent name (e.g., claude-code, cursor)
HOLLY_USER	User name
HOLLY_LLM	LLM name (e.g., claude-sonnet-4-6)

Auto-detected from CLAUDE_PROJECT_DIR, ANTHROPIC_MODEL, CURSOR_MODEL, etc.

Privacy

holly-db never phones home, never collects usage data, and never makes network calls except when downloading the embedding model via holly init. This is a permanent commitment.

License

MIT