Skip to main content

Analytics and insights

Updated on 3 November 2026.

MCP Gateway Insights shows how AI clients use your servers, what they touch, and why calls fail. Open Settings > AI Gateway > MCP Gateway, then select Overall analytics.

Choose All servers or narrow the server scope, then select All AI clients or a particular client. Period choices are 3 days, 7 days, 30 days, and 90 days, compared with the previous period of the same length. Panels with their own fixed window state it next to the figures.

Overview​

Start with tool calls, active people, failures and detections, then follow an attention item or chart into its details.

Topics & what AI touched summarizes topics classified from tool-call requests and responses, along with the coverage and the date classification began. A topic activity window is one person's activity in one six-hour UTC block. Classification coverage can differ from the full traffic count.

Use Needs attention, calls over time, top servers and active people to choose where to investigate. A monitored detection means the content was allowed through; it should not be counted as a prevented call.

Usage​

Use this tab to understand adoption and repeat use across people, servers and AI clients.

  • New vs returning counts a person as new when their first-ever call falls in the selected period.
  • Came back counts people active on two or more days in the period.
  • Servers per person, weekly activity and adoption by server show how broadly people use the gateway.
  • Client × server connects the AI clients people use with the servers they call.

Search the People table or filter it to New, Came back, Struggling, or DLP flags. Select a person or server for supporting details.

Data & security​

Review the resources tools touch, calls that change data, and flagged content moving in either direction.

The page distinguishes Read, Write, Execute, and Unclassified calls. The classification considers the tool and its operation arguments; the resource view shows calls that name a repository, project, site, channel or similar resource.

Use the access map, Changes made by AI, and tools that changed data to trace activity. In the finding views, switch between Responses → AI and Requests → tools. Review categories, tools returning flagged content, prompt-injection sources and affected people, then inspect the calls behind them. See Security guardrails to change protection.

Reliability​

Use this tab to understand failure causes and who can act on them.

Failures are grouped by causes such as Agent input, Upstream service, and Other. Select a cause to filter failure patterns, then review affected tools and people. Struggling identifies people whose failure rate is at least one in five across 20 or more calls.

Fixes you can make at the gateway can surface input-alias candidates. Review the evidence before accepting a mapping: similar argument names or success with an undeclared argument do not prove what the upstream tool expected. Reviewing analytics does not apply a fix. See Input aliases for that workflow.

Servers & tools​

Find unhealthy servers, unused tool definitions, and opportunities to shorten what AI clients receive.

Switch between Servers and Tools. Filter by Healthy, Has issues, Failing often, Low traffic, or Idle, and by whether calls came through OneMCP or a direct connection. Server health comes from calls in scope; Idle means no calls in 30 days.

Tool list savings shows estimated tokens kept out of AI turns by shorter tool lists, plus potential savings on other servers. Its figures use the stated 30-day window. Tokens returned to AI highlights large tool results; select a tool to inspect Overview, Output size, and Data protection.

Keep the two kinds of savings separate:

FigureWhat it measures
Tokens saved from response transformsSmaller tool results after token-saving strategies run.
Tool list savingsEstimated tool-description tokens not sent because a short list was used.

See Tool lists and Token saving for configuration and estimation details.

Open individual calls and settings​

Overall analytics opens Insights. For one server's call log, return to its card and select Inspect > Logs. The operational Activity view contains Tool calls, Interactions, and Findings. Select a call for its details.

Use the server card's Configure actions to change settings. The Insights tabs are for investigation; viewing a recommendation does not approve a tool, apply an input alias, or change policy.

On this page