toolguard: is this MCP tool safe?

Paste an MCP tool and the tools already registered on the server. toolguard-1b (a Gemma 3 1B fine-tune) says whether the tool tries to make an AI agent act against its user: hidden instructions, data exfiltration, manipulating other tools, impersonation, or a malicious update (rug pull). It runs entirely in your browser with llama.cpp (WebAssembly): nothing you paste leaves this page.

Registered tools (one per line: name: description) Candidate tool name Candidate description Parameters: JSON list of {name, description, required} or an MCP inputSchema Previous approved description (optional, to check an update)

Examples:

Results on toolguard-bench (Q4_K_M): 86.8% of poisoned tools flagged, 1.3% false alarms on 305 real tools. It is a fast first filter, not a guarantee. Speed depends on your device; on a 4-core server CPU a check takes about 2 s.