Kernal

Open-source knowledge graph for professionals. Auto-extracts entities and relationships from natural conversation via MCP.

Talk to Claude naturally about your meetings, calls, and interactions. Kernal stores people, organizations, topics, and relationships — building a knowledge graph you own.

What's Included (Open Source)

Everything you need to run Kernal locally on your own machine:

• 13 MCP tools — ingestion, CRUD, query, corrections (see full list below)

• SQLite database — local-first, your data never leaves your machine

• LLM-driven extraction — Claude reads your text, decides what to extract, and calls structured write tools

• Entity resolution — fuzzy matching + Levenshtein distance prevents duplicates

• CLI — init, serve, status, export

• Cloud server — Express.js with API key auth, rate limiting, CORS, session management

• Dashboard — React app with network graph, timeline, action items, overview

• 50 tests — comprehensive test suite

This is a fully functional knowledge graph you can run yourself, for free, forever.

Related MCP server: Beever Atlas

What Andes Provides (Managed Service)

For teams and professionals who want more, Andes offers:

• Cloud hosting — access your knowledge graph from any device, no self-hosting

• Dashboard — hosted interactive visualizations powered by your data

• Multi-user — team features, shared knowledge bases, role-based access

• Onboarding & support — we set it up for you and help your team get value from day one

• Industry workflows — pre-built patterns for executive search, consulting, professional services

The open-source core is the engine. Andes wraps it with infrastructure, UX, and support.

Quick Start

npx kernal-mcp init

This creates a SQLite database at ~/.kernal/kernal.db and prints the config to add to Claude Desktop.

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "kernal": {
      "command": "npx",
      "args": ["-y", "kernal-mcp", "serve"]
    }
  }
}

Restart Claude Desktop. Then talk naturally:

"I had lunch with Jonas Lindberg from Nordvik Energy today. He's their VP of Digital. We discussed their cloud migration — targeting Q3."

Claude extracts Jonas, Nordvik Energy, the cloud migration topic, and stores them via Kernal's write tools. Then ask:

• "What do I know about Nordvik Energy?" → Full briefing with people, interactions, topics

• "Who should I follow up with?" → Open action items with owners and due dates

• "Show me everyone at Nordvik Energy" → Contact list filtered by organization

How It Works

Kernal uses an LLM-driven extraction pattern:

1. You tell Claude about a meeting, call, or interaction

2. Claude calls kernal_remember with the raw text

3. Kernal stores the text as a note and returns extraction instructions + existing entities (for dedup)

4. Claude reads the text intelligently and calls structured write tools (kernal_add_person, kernal_add_org, kernal_add_activity, etc.)

5. Each write goes through entity resolution to prevent duplicates

6. The LLM makes all extraction decisions — no regex guessing

The MCP server is a clean data store. The LLM is the brain.

MCP Tools

Ingestion (write)

Query (read)

Corrections

What Gets Stored

From a single paragraph like "Had coffee with Sofia Andersen from Arctura Tech. She's their VP of Sales. We discussed their expansion into APAC. I need to send her the partner proposal by Friday.", Claude will call:

• kernal_add_person — Sofia Andersen, VP of Sales, at Arctura Tech

• kernal_add_org — Arctura Tech

• kernal_add_activity — Coffee meeting, today, participants: [Sofia Andersen], orgs: [Arctura Tech]

• kernal_add_action — "Send partner proposal to Sofia", due Friday, owner: Sofia Andersen

• kernal_link — Sofia → works_at → Arctura Tech

Each call is a deliberate, structured decision by the LLM — not a regex guess.

CLI Commands

kernal init      Create database + print Claude Desktop config
kernal serve     Start MCP server (stdio transport)
kernal status    Show database stats
kernal export    Export database to a file
kernal help      Show help

Dashboard

The repo includes a React dashboard (dashboard/) with four views:

• Overview — entity counts, most connected people, activity breakdown

• Network — interactive force-directed graph (people + organizations)

• Timeline — chronological activity feed with participants and summaries

• Actions — follow-ups grouped by urgency (overdue, this week, upcoming)

Natural language command bar routes queries to views ("Show me my network" → graph).

# Start the cloud API server
KERNAL_API_KEY=your-key KERNAL_DB_PATH=~/.kernal/kernal.db npm run cloud

# Start the dashboard (separate terminal)
cd dashboard && npm run dev

Data Model

Kernal stores 6 entity types connected by a generic relationship graph:

People ←→ Organizations
  ↕           ↕
Activities ←→ Topics
  ↕
Actions ←→ Notes

All entities can link to any other entity via the relationships table, enabling queries like:

• "Who has Sofia met with?" (person → activities → other people)

• "What topics come up with Nordvik Energy?" (org → people → activities → topics)

• "What's the connection between Jonas and Arctura Tech?" (path through graph)

Security

• All SQL queries use parameterized statements (no injection risk)

• API key auth with constant-time comparison (crypto.timingSafeEqual)

• CORS restricted to configured origins

• Rate limiting (120 req/min per IP, configurable)

• MCP session timeout (30 min idle eviction)

• No secrets in code — all config via environment variables

• React dashboard auto-escapes all rendered data (no XSS)

Development

git clone https://github.com/pintomatic/kernal.git
cd kernal
npm install
npm run build
npm test        # 50 tests

Self-Hosting the Cloud Server

KERNAL_API_KEY=your-secret KERNAL_DB_PATH=~/.kernal/kernal.db npm run cloud

A Dockerfile is included. Environment variables:

Seed Demo Data

npx tsx scripts/seed-demo.ts

Creates 12 contacts, 18 orgs, 19 activities with 123 relationships — a realistic professional services scenario.

License

MIT