Restrict the tools your AI agents can call.

One allowlist per agent, enforced at the gateway. Tools that are not on the list are removed before the model sees them, and disallowed calls that come back anyway are blocked.

X-Grepture-Label: agent:support-botPOST /v1/chat/completions

Request: tool definitions

  • search_kbkept
  • create_ticketkept
  • get_order_statuskept
  • refund_orderremoved
  • delete_accountremoved
tool_choice"refund_order"→"auto"reset

Response: tool call

Model returned a tool it was not given

delete_account({"user_id": "u_4821"})

On a disallowed call
403Response rejected. Your app never acts on the call.

What it does

You list the tools an agent may call. Grepture enforces that list on every request, whatever your application code registers. It works with OpenAI Chat Completions, the OpenAI Responses API and Anthropic Messages.

How it works

Scope a rule to one agent with the X-Grepture-Label header, for example agent:support-bot. On the request side, the gateway removes every tool definition that is not allowed and resets a forced tool_choice on a removed tool to auto. On the response side, a disallowed tool call is either blocked with a 403 or stripped. Response checks apply to non-streaming responses; request-side stripping covers streaming too.

Least privilege for agents.

  • Prevention by omission

    Disallowed tools are removed before the model sees them, so there is nothing for an injection to hijack.

  • A backstop on the way back

    A disallowed call that comes back anyway is blocked with a 403 or stripped.

  • One allowlist per agent

    Scope rules with X-Grepture-Label, so each agent gets only the tools it needs.

  • Built from real traffic

    The rule editor suggests the tools your agents have actually called.

What's included

  • One allowlist per agent, scoped with X-Grepture-Label
  • Disallowed tool definitions are removed before the request is forwarded
  • A forced tool_choice on a removed tool is reset to auto
  • Disallowed calls in the response are blocked (403 by default) or stripped
  • OpenAI Chat Completions, OpenAI Responses API and Anthropic Messages
  • The rule editor suggests tools your agents have called, from your traces
  • Request-side stripping also protects streaming responses
  • Set up under Guardrails → Rules → Restrict Tools, live within seconds

Send your first clean request in five minutes.

Free for 1,000 requests a month. No credit card, no sales call.