AgentGate Adds an Authorization Gateway for Autonomous AI Agent Tools

Summary

AgentGate is an open-source, out-of-process authorization proxy and security gateway for autonomous AI agent tool execution. It intercepts tool calls before they reach production infrastructure and combines a sub-40-millisecond neural classification step with deterministic access-control policies. The project is designed around the risk that untrusted content in an agent's context can influence a destructive action even when system instructions remain unchanged. Its decision service runs a fine-tuned Qwen 0.8B classification model through Unsloth, with a calibrated CPU stub available for environments that cannot run the model. Four classification heads assess whether an action reflects untrusted prompt influence, whether it aligns with the declared session goal, how reversible or costly it is, and whether it sends data externally. AgentGate can return allow, deny, or ask; the last state places the decision in a PostgreSQL-backed operator review queue. The proxy also stores audit logs and decision traces, exposes metrics, and provides a React and Vite dashboard with live events, approvals, and policy replay. The repository includes integrations and examples for MCP, TypeScript, LangChain, CrewAI, Python, and REST, plus policy, deployment, architecture, and API documentation. Its published comparison reports 38 ms p95 latency and self-hosted zero cost per 1,000 calls, versus 1,150 ms for Claude 3.5 Haiku and 1,420 ms for GPT-4o-mini in the listed LLM-as-a-judge tests; those alternatives are described as jailbreakable in the project’s benchmark. A concurrent test command is reported to reach 262 requests per second with 53 ms p95 latency. The included security verification lists six scenarios covering indirect prompt injection, exfiltration, destructive file changes, irreversible shell actions, and benign operations, with all expected verdicts reported as passing. The project is MIT-licensed and can be started locally with Docker Compose, while a live playground and model resources are linked from the repository.