· announcement / mcp / security
Introducing AgentGate: The Authorization Layer for AI Agents
AI coding agents stopped being chatbots. Claude Code, Cursor, Cline, Codex and countless custom agents now read files, run shells, query databases and call APIs through the Model Context Protocol (MCP). That is enormously useful — and it creates a problem every security team recognizes:
Software that can act now needs authorization.
We built AgentGate to be that layer.
What AgentGate is
AgentGate is a runtime authorization and security control plane for AI agents. It sits between your agents (any MCP-compatible client) and the MCP servers they use. Every tool call is intercepted before execution and evaluated against policy:
- ALLOW — the call matches an explicit policy and passes risk checks.
- DENY — no matching policy exists (the default), or the call is too risky.
- APPROVAL_REQUIRED — a human reviews the call and approves or rejects it.
Every decision is written to an append-only audit trail: which agent, which tool, which arguments, which policy decided, and who approved it.
Why a gateway, not a plugin
Authorization has to live where it cannot be bypassed by the thing being authorized. A plugin inside the agent depends on the agent behaving; a gateway in the network path does not. AgentGate runs as its own service, in front of your MCP servers, with:
- a default-deny policy engine (a call with no matching policy never passes),
- risk scoring on top of binary policy, so you can grade calls instead of only gating them,
- human-in-the-loop approvals with one-time tokens, expiry and replay protection,
- secret detection that keeps raw keys out of logs and audit records,
- fail-closed behavior — if the policy engine or database errors, the call is blocked.
Available now
AgentGate enforces deterministic policy before forwarding allowed calls to registered upstream MCP servers. Teams can create a workspace, register an agent and MCP server, simulate policy, review approvals and export audit evidence today. The changelog tracks product releases.
If your team runs AI agents against real infrastructure, start with a 14-day technical trial and a disposable MCP workflow. Teams planning a private or self-hosted rollout can talk to enterprise.
The question is no longer whether agents will call tools on your systems. It’s whether someone will be authorizing those calls when they do.