Skip to main content

Tool controls

Updated on 3 November 2026.

Allow or deny tool calls by tool name, declared annotations, wire type or argument values, independently of what data they carry.

Where it is configured​

Tool controls are a Policy Engine feature. There is no app setting or Configure tab for them, so nothing freezes when you use them. Open Govern > Policy Engine > LLM Gateway > Controls > Tool Controls. Start from a recipe or the Tool calls add menu, choose Applies to and the outcome, then Save to draft. Edits join the shared draft and take effect once you review and publish a revision.

Sections​

SectionWhat it doesAdd button
Tool callsPer-call allow or deny rules.Section add menu
Argument & result protectionLists the Data & Adversarial Risks rules scoped to the tool boundary. Edit them in that control.Open Data & Adversarial Risks

Tool calls applies on assistants, bedrock, chat, copilot, embeddings, rerank, responses, sdk_check, stt, text, tts and vertex.

Every tool call in a request is judged on its own, but one denied call rejects the whole request. The engine never strips a single call. Per-call verdicts stay visible in attribution.

Tool rule settings​

SettingOptionsDefaultNotes
ToolTool names or patterns (delete_*), or Any toolEmptyNames and patterns are alternatives. Leave empty or tick Any tool to judge every call by traits or arguments.
By traitDestructive, Open world, Not read-only, Read-only, IdempotentNoneDeclared MCP-style annotations read from the call. Match any trait (default) or all traits.
Tool typeWire type, for example functionEmptyPatterns allowed.
When argumentsArgument path + is, is not, one of, contains, starts with, ends with, matches *wildcard*, exists, is emptyNonePaths are suggested per tool from observed calls; nested paths like options.visibility work. Text conditions ignore case. Values are never stored. Match all (default) or any.
DecisionAllow, DenyDenyA deny rejects the whole request that carries the call.
SeverityNot set, Very low, Low, Medium, High, Critical, Very criticalNot setReported only.

Risk and tags are not computed by the gateway, so the quick rule does not offer them. Use Add configuration for result-field conditions on the response stage or combinations the quick rule cannot express.

Scenarios​

  • Block destructive tools for most people. Deny calls whose trait is Destructive for Everyone, then allow them at a higher priority for one Smart group.
  • Block one risky argument value. Deny create_repository when visibility is public:
deny_public_repository_creationrequest

runs on requestpriority 950

WhenTool nameiscreate_repository
andTool arguments . visibilityispublic
Then
Tool call accessdeny
Risk levelhigh
  • Pattern rules. Deny every delete_* tool for one Application or Requested model.

To scan tool arguments for secrets or PII instead, add a data rule with Scan tool-call arguments on Data & Adversarial Risks. Only Monitor and Block apply there, because redaction cannot preserve a call.

Scoping and precedence​

  • Applies to: Everyone, People, Smart group, Application, App tag, Requested model, Provider, API surface, Environment, Tool, Source network, or Except....
  • Highest priority wins for each call; Deny wins equal-priority ties. Without a matching rule, tool calls pass this control and still follow the data-risk checks.

Test before you publish​

Open Test > Try a situation in the LLM Gateway Policy Engine. Describe a matching request, run the test and review the rules that decide its settings. Repeat with a caller or tool that should remain allowed. Use Replay activity to check recent recorded traffic, then Review & publish when the draft is ready.

Use the Policies tab for the complete rule list and Edit technical source for options the guided form cannot express. History shows published versions and their changes.