Solarium

A knowledge base MCP server backed by Qdrant vector database with local embeddings (nomic-embed-text-v1.5 via HuggingFace Transformers + ONNX Runtime).

Solarium runs as a local MCP server process — embeddings are computed on your machine, vectors are stored in a Qdrant instance of your choice (local or cloud).

Setup

Prerequisites

• Node.js 20+

• npm

Install dependencies

cd solarium
npm install

Build and link

npm run build
npm link

This builds the server and creates a global solarium command on your PATH.

Run the solarium binary

solarium

Pre-download the embedding model (optional but recommended)

The model (~130MB) is downloaded automatically on first use, but you can cache it ahead of time:

./bin/download-model.sh

Configure your Qdrant instance

Set your Qdrant API key in your shell profile (~/.zshrc or ~/.bashrc):

# Qdrant Cloud (Solarium)
export QDRANT_API_KEY=<your-qdrant-api-key>

Then reload your shell: source ~/.zshrc

Add to your Claude Code MCP config

Add the following to .mcp.json in any repo where you want Solarium available:

{
  "mcpServers": {
    "solarium": {
      "command": "solarium",
      "env": {
        "QDRANT_URL": "https://your-cluster.cloud.qdrant.io:6333"
      }
    }
  }
}

QDRANT_API_KEY is read from your shell environment automatically.

Verify

In Claude Code, run /mcp to confirm Solarium is connected, then try:

Search the knowledge base for "deployment architecture"

Related MCP server: team-docs-mcp

Available Tools

Dashboard

Solarium includes a local web dashboard for browsing documents, semantic search, tag management, and analytics.

Build and run the dashboard

Reload your shell first.

npm run build:dashboard
solarium-dashboard

Then open http://localhost:3333 in your browser.

The dashboard requires the same QDRANT_URL and QDRANT_API_KEY environment variables as the MCP server. Set DASHBOARD_PORT to change the port (default: 3333).

Views

• Overview — collection stats (document/vector/point counts), recent documents, tag cloud

• Documents — paginated document list with tag filtering, full document detail with rendered markdown and chunk visualization, inline tag management (add/remove)

• Search — semantic search with relevance scores, snippets, and tag filtering

• Analytics — documents over time, tag distribution, chunk size distribution (Vega-Lite charts), collection stats

Architecture

The dashboard runs as a separate process from the MCP server (which uses stdio). Both share the same source code — Qdrant client, tools, config, and embedding pipeline.

• Server-driven rendering — HTML is rendered on the server from ClojureScript hiccup, streamed to the browser

• SSE + Idiomorph — live updates via Server-Sent Events with DOM morphing (polls Qdrant every 60s)

• Tailwind CSS + @tailwindcss/typography — styling with the prose class for markdown content

• Vega-Lite — interactive charts, lazy-loaded from CDN only on the analytics page

• No frontend framework — no React, no client-side state management; ~30 lines of client JS for SSE wiring

Configuration

All configuration is via environment variables:

Development

# Watch mode — MCP server
npm run watch

# Watch mode — dashboard
npm run watch:dashboard

# Watch both
npm run watch:all

# Run tests
npm test

Built with shadow-cljs (ClojureScript).