MCP server for Glassbox — mechanistic interpretability + EU AI Act Annex IV compliance tools for Claude and any MCP-compatible client
Wellknown found it in public sources; nobody has proven control of it yet. Claiming takes one click if the repository is under your GitHub account, or a small file on your domain otherwise. Verified owners get the badge, 15-minute checks, status alerts, edits that outrank crawled data, and a ranking boost.
Agents can do it too: POST https://wellknown.network/api/v1/claims with {"agent":"glassbox-mcp","method":"well_known_file"} — machine-readable steps at claim.json, guide at /docs/claim.
Everything here was measured by our prober or read from a registry. Nothing is self-reported.
Attributed to the source that supplied each field. Treated as claims, not facts.
# glassbox-mcp MCP server for [Glassbox](https://project-gu05p.vercel.app) — mechanistic interpretability + EU AI Act Annex IV compliance tools, exposed via the [Model Context Protocol](https://modelcontextprotocol.io). Connect Claude (or any MCP-compatible client) to Glassbox with a single pip install. No cloning, no manual server setup. ## Install ```bash 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): ```json { "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](https://arxiv.org/abs/2603.09988) - PyPI (core library): […
Mapped onto the structured taxonomy from declared text and observed tool names. Confidence shown for derived entries.
Every source is kept verbatim. Field changes are logged as events.