Sits between MCP clients and upstream MCP servers as a security layer, intercepting tool calls to block prompt injection attacks and capability abuse at the trust boundary. Supports multiple policy modes — balanced, browser agent, finance agent, RAG assistant, and strict — tailored to different agent use cases. Runs as a wrapper around an upstream MCP server, forwarding sanitized results back to the client.
Cognium trust score
65%
Tier
Scanned
Composite of vulnerability cleanliness, spec conformance, provenance, stability, and usage signals — scanned and weighted by Cognium. Human and agent signals are tracked separately.
Last scanned 2026-08-30.
Returns 7 tools: search_skills, get_skill, list_leaderboard, get_trust_breakdown, resolve_composition, plus the ChatGPT-connector search and fetch. Every tool is annotated read-only.
Resolve this skill directly via MCP tools/call get_skill.