Agent Approval plugin
The Agent Approval plugin requires a human approval token on requests that match its conditions, such as a method, a path, or an MCP tool. An agent sends the token in a header, or, on an MCP tools/call, in an approval parameter. Without a valid token, the request is refused. Only agents are asked unless apply_to_humans is on.
A caller is an agent when JWT Auth identified it and its token marks it as an agent.
Where it runs
Configuration
Methods that need approval, in any letter case. The console labels this Require methods.
Exact paths that need approval.
Path prefixes that need approval. The console labels this Require path prefixes.
A caller holding any of these scopes needs approval. The console labels this Require scopes.
MCP tool name patterns that need approval. A pattern ending in * matches by prefix, and one starting with * matches by suffix.
The header that carries the token.
How to check the token. static checks it against static_tokens. A data plane set up with an identity service defaults to a different validator, so always set validator to static and fill in static_tokens. The console labels the blank option “default (iam)”.
Maps an approver name to sha256: followed by the full 64-character hexadecimal SHA-256 of that approver’s secret. The gateway checks the stored value only as a prefix, so a shorter or empty value approves other secrets.
Also asks users and callers with no agent identity for approval.
With no require_for condition, nothing needs approval. A request needs approval when any one condition holds.
With the static validator, the token has the form <approver>:<secret>. A missing token gets “human approval required”, and an invalid one gets “approval token rejected”. The token stays on the request and is forwarded to the upstream.

