A knowledge graph server for LLM agents, built on Bun with SQLite and optional Dolt.

gk is an MCP server that gives agents tools to build, search, and analyze knowledge graphs. The schema is fully dynamic — the agent decides what entity types, relationship types, and properties fit the domain. One database per project.

27 tools across four tiers:

Built-in domain guidance — Four guides ship as both MCP prompts and resources, covering extraction, pyramid observations, querying, and review.

Temporal dynamics — Hebbian strengthening on access (stability grows), Ebbinghaus decay over time (unused knowledge fades in search rankings). Overview-tier knowledge is architecturally durable; details are ephemeral.

• Bun — Runtime, test runner, package manager
• SQLite via bun:sqlite — Embedded database with FTS5 full-text search
• sqlite-vec — Vector similarity search for semantic embeddings
• Ollama (optional) — Local embeddings via nomic-embed-text for semantic search
• Dolt (optional) — MySQL-compatible database with git-like versioning, via Bun.SQL
• MCP SDK — Model Context Protocol server

# Clone and install
git clone https://github.com/yourusername/gk.git
cd gk && bun install

# Initialize a database
bun run . init

# Run tests
bun test

Add to your .mcp.json:

{
  "mcpServers": {
    "gk": {
      "type": "stdio",
      "command": "bun",
      "args": ["run", "--bun", "/path/to/gk"],
      "env": {
        "GK_DB_PATH": "/path/to/your/project/knowledge.db"
      }
    }
  }
}

Environment variables or gk.yml:

bun test              # Run tests (119 pass, 7 Dolt tests skip without GK_DOLT_HOST)
bun run check         # Biome lint + format check

MIT