What it is

Toolfence is the security and governance layer above OAuth, between AI agents and the MCP servers they call. It ships as two products:

Scanner (toolfence, open source, Apache 2.0) — npx toolfence <url> flags tool poisoning, prompt injection, behavioral drift, and scope/cost issues in an MCP server before your agents connect to it.
Gateway (proprietary, hosted) — a runtime proxy that sanitizes tool responses, enforces allow-lists and rate/cost ceilings, and audits every call between agents and MCP servers.

Why now

MCP standardized the substrate for agent tool calls (7,300+ servers and growing). OWASP named tool poisoning and prompt injection via tools in the Top 10 for LLMs. There's no managed security layer between agents and their tools — just raw connections, no audit trail, no sanitization, no scope control.

Current status

Scanner live on npm. Gateway prototype built. Landing page live.

Stack

TypeScript, npm.