JetStream Security, a Santa Clara-based startup building governance tools for enterprise AI agents, on Tuesday introduced Clearance, a system designed to evaluate every action an AI agent attempts and decide whether it should be allowed to run before it executes.
Clearance sits at the point where an agent calls a tool through the Model Context Protocol, the increasingly common standard connecting AI systems to outside software. When an agent requests an action, Clearance checks it against what the company calls an "AI Blueprint" — a record of which models, tools, datasets and identities a given agentic system is authorized to use — down to the level of individual parameters. A team could, for instance, let an agent read from a database while blocking it from deleting records, without writing custom logic for every tool.
The pitch is pre-emptive rather than forensic. Most existing AI gateways log what an agent did after the fact, giving security teams a record to investigate once something has gone wrong. JetStream co-founder and chief executive Raj Rajamani said that approach falls short as companies move from AI pilots to agents that act with real consequences.
A gateway can tell you where a request went. The question our customers keep asking is whether the agent should have sent it at all.
Raj Rajamani, co-founder and chief executive, JetStream Security
Co-founder and Chief Architect Venu Vissamsetty said the gap between deployment and oversight is widening as agent fleets grow. JetStream, founded by veterans of CrowdStrike, SentinelOne, Cohesity and Dazz, raised $34 million in seed funding earlier this year from Redpoint Ventures and the CrowdStrike Falcon Fund, alongside angel investors including CrowdStrike co-founder George Kurtz and Wiz co-founder Assaf Rappaport.
Clearance is aimed at enterprises the company says are preparing to run hundreds of thousands of AI agents across customer service, IT and back-office functions — a scale at which manually reviewing every permission becomes impractical. JetStream plans to demonstrate the tool at CrowdStrike's Fal.Con conference this month, with general availability expected this fall.
The launch adds to a fast-growing category of AI "control plane" products competing to answer a question enterprises are asking with increasing urgency: not whether to deploy autonomous agents, but how to keep one from taking an action nobody approved.
