Inline Gateway
Inside the gateway
Fendray sits in the request path and acts before each tool call executes — allow, block, or step up, in real time. Every call is logged in an immutable, queryable audit trail, attributable to an agent and its owner. Here's what that looks like.
Console shown with sample data.
Every rule is readable
Policies are version-controlled and published as code — allowed tools, step-up thresholds, PII detection, destination regions, injection detection, all explicit. The gateway enforces a published policy within ~30 seconds. When an auditor asks why a call was stopped, the answer is a line of policy, not a model's guess.
Every agent, governed
Across LangChain, CrewAI, AutoGen, and MCP — every agent registered, with its owner, team, active policy, and status in one inventory. No registered agent operating outside policy.
A human in the loop, only when it matters
When an action crosses a threshold you set — a payment above a limit, a sensitive write — the gateway stops it until a human approves. The reviewer sees the agent, the tool, the action, and the amount, then approves or rejects. Everything under the threshold proceeds untouched.