Context45

All the docs. None of the bloat.

Context45 is a minimal MCP server that gives AI assistants only the essential documentation context. Instead of flooding your context window with 5,000+ tokens of docs (like Context7), Context45 returns 100-300 tokens of precisely what you need.

The Problem

Tools like Context7 inject documentation into AI assistants, but they return way too much:

| | Context7 | Context45 | |-|----------|---------------| | Tokens per query | 5,000-6,200 | 100-300 | | Input multiplier | 155-215x | 3-10x |

That's a 20x reduction in context usage while still giving you the right answer.

Quick Start

Claude Code

claude mcp add context45 -- npx context45-mcp

Then in your prompts, add "use context45" to fetch relevant docs.

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "context45": {
      "command": "npx",
      "args": ["context45-mcp"]
    }
  }
}

Available Libraries

| Library | Status | |---------|--------| | Claude API (Anthropic) | ✅ Available | | OpenAI API | ✅ Available |

How It Works

1. Your AI assistant calls resolve-library to find the library
2. Then calls get-docs with your query (e.g., "how to stream responses")
3. Context45 does a semantic search over pre-compressed documentation
4. Returns only the most relevant 2-3 chunks (~100-300 tokens)

The documentation is manually curated and compressed — we strip tutorials, marketing, migration guides, and niche features. Only what 95% of developers actually need day-to-day survives.

MCP Tools

resolve-library

Find a library by name.

Input: { "query": "claude" }
Output: Found 1 matching library:
  - Claude API (id: claude): Anthropic Claude API reference...

get-docs

Get minimal documentation for a specific query.

Input: { "libraryId": "claude", "query": "streaming" }
Output: ~146 tokens of streaming docs with code examples

Parameters:

• libraryId (required): Library ID from resolve-library
• query (required): What you need docs for
• maxResults (optional): Number of chunks to return (default: 3, max: 10)

Development

Prerequisites

• Node.js >= 18
• Upstash Vector index (with bge-m3 embedding model)

Setup

git clone https://github.com/arbab-ml/context45.git
cd context45
npm install
cp .env.example .env
# Add your Upstash Vector credentials to .env

Adding Documentation

1. Place a curated _processed.md file in docs/<library>/
2. Run the pipeline:

npm run pipeline

This chunks the markdown and uploads it to Upstash Vector.

Standalone Testing

Test the MCP server directly without an AI assistant:

# Start the server and send a get-docs request
printf '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"0.1.0"}}}\n{"jsonrpc":"2.0","method":"notifications/initialized"}\n{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"get-docs","arguments":{"libraryId":"claude","query":"streaming"}}}\n' | npx tsx src/server/index.ts 2>/dev/null

This sends JSON-RPC messages over stdin (the MCP stdio transport) and prints results to stdout.

Development

See README-dev.md for the full development guide — adding libraries, running the pipeline, cleaning vectors, etc.

License

MIT