Needle MCP Server

A production-ready Model Context Protocol server that runs the Cactus Compute Needle model locally for structured extraction, classification, summarization, and context pruning. Keeps your data on device.

About

Cactus Compute Needle is a tiny, edge-first foundation model built for tool calling and structured extraction, not chat or facts. Needle 2 is a 45M-parameter model shipped as a 14MB binary that runs in about 28MB of RAM. Its architecture favors cross-attention over memorization-heavy feed-forward layers, which is what makes it practical for on-device agents.

The model lives at github.com/cactus-compute/needle and the wheels are published on Hugging Face. This MCP server downloads the appropriate engine for your platform on first run and caches it under ~/.cache/needle, then exposes it through six read-only tools.

Related MCP server: ClawMem MCP Server

Features

• Extract JSON-structured data from any text without writing a prompt.

• Automatically drops unnecessary context to stay inside the model context window.

• Chunk and search large documents locally before sending them to a large language model.

• A JSON repair step guarantees parsable results.

• SQLite cache returns identical requests instantly.

Needle MCP exposes a local Model Context Protocol server that runs the Cactus Needle model on-device. It turns structured-extraction work — JSON extraction, classification, summarization, context pruning — into callable tools, so agents and CLIs can offload that work without a remote API.

How it works

The server boots a single process over stdio. On startup it downloads the Needle engine wheel for the host platform (or reuses a cached copy under ~/.cache/needle), then initializes an MCP Server. Every tool call is a single pass through the local model. Identical requests — same tool name and arguments — are read back from a SQLite cache at ~/.cache/needle/mcp_cache.db, so repeats return instantly and never re-hit the model.

Inputs that exceed the context window are handled in-process: large texts are chunked and trimmed to the relevant slices before being handed to the model, which keeps latencies low and the model honest.

Installation

uv tool install git+https://github.com/yugaaank/needle-mcp

Usage

Claude Code

claude mcp add needle -- needle-mcp

Cursor

cursor mcp add needle -- needle-mcp

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json, Windows: %APPDATA%\Claude\claude_desktop_config.json):

{
  "mcpServers": {
    "needle": {
      "command": "needle-mcp"
    }
  }
}

Restart Claude Desktop for the change to take effect.

Antigravity CLI

agy mcp add needle -- needle-mcp

OpenCode

Add the server to your opencode.json under the mcp key:

{
  "mcp": {
    "needle": {
      "type": "local",
      "command": ["needle-mcp"],
      "enabled": true,
      "timeout": 30000
    }
  }
}

Oh My Pi (OMP)

Add the server to an OMP-native MCP config file:

• Project-scoped: .omp/mcp.json

• User-wide: ~/.omp/agent/mcp.json

{
  "$schema": "https://raw.githubusercontent.com/can1357/oh-my-pi/main/packages/coding-agent/src/config/mcp-schema.json",
  "mcpServers": {
    "needle": {
      "command": "needle-mcp"
    }
  }
}

Or use the interactive wizard in a running OMP session, then reload:

/mcp add
/mcp reload

CLI

needle-mcp

Tools

All tools stream structured output back as JSON. Results for identical text + arguments are cached locally, so the second call is a cache hit.

License

MIT – see the LICENSE file for details.