glassbox-mcp 4.2.4

MCP server for Glassbox — mechanistic interpretability + EU AI Act Annex IV compliance tools for Claude and any MCP-compatible client

glassbox-mcp

MCP server for Glassbox — mechanistic interpretability + EU AI Act Annex IV compliance tools, exposed via the Model Context Protocol.

Connect Claude (or any MCP-compatible client) to Glassbox with a single pip install. No cloning, no manual server setup.

Install

pip install glassbox-mcp

Connect to Claude Desktop

Add to ~/.claude/claude_desktop_config.json (macOS/Linux) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

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

Restart Claude Desktop. The Glassbox tools appear immediately.

Tools

Tool	What it does
glassbox_circuit_discovery	Attribution patching — find which attention heads causally drive a prediction
glassbox_faithfulness_metrics	Compute sufficiency, comprehensiveness, F1, and EU AI Act explainability grade (A–D)
glassbox_compliance_report	Generate a full EU AI Act Annex IV evidence package (all 9 sections)
glassbox_attention_patterns	Get attention weight heatmap for any layer/head
glassbox_logit_lens	Layer-by-layer residual stream projection — see how predictions build up

Supported Models

• gpt2, gpt2-medium, gpt2-large
• EleutherAI/pythia-70m, pythia-160m, pythia-410m

Example usage in Claude

Once connected, ask Claude things like:

"Use Glassbox to analyse which attention heads drive GPT-2's prediction for 'The Eiffel Tower is in' → ' Paris'"

"Generate an EU AI Act Annex IV compliance report for GPT-2 with provider name Acme Corp"

"Show me the logit lens for GPT-2 on this sentence"

Requirements

• Python ≥ 3.10
• ~3 GB disk (PyTorch + TransformerLens model weights download on first use)

Version

4.2.4 — tracks glassbox-mech-interp v4.2.4

• Paper: arXiv 2603.09988
• PyPI (core library): glassbox-mech-interp
• Website: project-gu05p.vercel.app
• GitHub: designer-coderajay/Glassbox-AI-2.0