Tool visibility
Updated on 3 November 2026.
Choose what AI agents can discover and call on an MCP server. Classic settings use per-tool switches; the Policy Engine can decide listing, calling, and returning results separately.
Choose the instructions for the mode that governs your MCP Gateway. See Switching from classic settings for what happens to stored settings when the engine is on.
- Classic settings
- Policy Engine
Configure it on the server
Open Configure > Tools on the server card. The section subtitle shows how many tools are exposed, for example 12 of 15 exposed.
Tool groups
Tools are grouped by what they do to the upstream system.
The Exposed to agents bar at the top shows how many tools in each group are enabled, for example Write 3 / 4.
Per-tool controls
Changes are drafts. Click Save settings in the footer to apply them.
Find a tool
Refresh the tool list
Click Refresh tools to fetch the server's current tool list. If the list is empty, the gateway hasn't fetched the tools yet; refresh, or connect OAuth first if the server needs it.
The Upstream tool changes card above the tool list watches for tools the server adds, removes or changes. See Tool change watch.
Per group and per user
To enable a tool, or require confirmation, only for some people, add a rule in Group and user rules. A rule can set each tool's Enabled, Confirmation and Justification to Inherit, On or Off for one smart group or one user.
Going further with the Policy Engine
Open Govern > Policy Engine > MCP Gateway > Controls > Invocation & Discovery. This single panel replaces the separate Discovery Visibility and Invocation entry points. Its sections are Tool calls, Resource reads, and Prompt fetches.
Each configuration can decide one or more access steps:
The What happens at each step table shows the steps each rule sets. Unset means that rule does not decide the step. A step is denied unless a matching rule allows it. For each decision, the highest numeric priority wins and Deny wins ties.
Allow or block tools
- Choose a recipe such as Allow tools on a server, Allow selected tools, Block a tool, or Block destructive tools.
- Set Applies to and choose the server, tools, and caller conditions you need.
- Check the steps the rule changes and its Reads as summary.
- Save to draft, test the situation in Test, and review and publish.
Hide from listings only changes discovery, not permission to call by name. An AI app may cache an old list, so use Block a tool when both listing and calling must stop. Argument-dependent recipes refuse matching calls; arguments are not available when a tool list is built.
Use Read-only tools, confirm the rest or Allow and require confirmation when tools should remain callable with user confirmation. The separate Human Confirmation control manages confirmation rules and exemptions.
Resources, prompts and results
Resource recipes can match URI, template, name or MIME type; prompt recipes can match a prompt name. Result recipes can withhold unsuccessful or oversized results. A condition on the returned result belongs to the result step, rather than discovery.
Open Activity to investigate observed matches. Use Test > Try a situation or Replay activity to check the candidate rules before publishing.
Related
- Human approval - what approvers see in each AI client.
- Tool change watch - review upstream tool changes before agents see them.
- Input aliases - translate mismatched tool inputs from AI clients.
- Security guardrails - scan tool inputs and results.
- Policy Engine overview - how cards, stages and priorities work.