Skip to main content

Analytics and insights

Updated on 3 November 2026.

LLM Gateway Insights shows how applications use models, where time and money go, why requests fail, and what guardrails find. Open Settings > AI Gateway > LLM Gateway, then select Overall analytics.

Choose All apps or narrow the application scope. Select 24 hours, 3 days, 7 days, 14 days, or 30 days; the page compares that window with the previous period of the same length. On tabs that offer Model requests and Guardrail checks, choose the traffic you want to inspect before comparing results.

Overview​

Start here to find a change worth investigating. Worth a look highlights signals, while the charts connect usage, caching, answer endings, model choices, failures and estimated cost.

  • What your apps do groups apps by their observed request patterns.
  • Prompt caching by app helps you find traffic that benefits from provider prompt caching.
  • What changed in your apps links configuration changes to the apps affected.
  • How answers end distinguishes finished answers, tool calls, output-limit cutoffs and other endings.
  • Models and routing, Calling clients, Where the money goes, and Failed requests show where to investigate next.

Select Explore, an app, or a chart entry to open its supporting details. The overview uses hashes, lengths and counts from request bodies; it does not display prompt or answer text.

Agents & tools​

Use this tab to understand runs, context growth, tool use and spend. All agents provides the overview; Tools & parameters focuses on the tools applications offer and call.

An agent here is one application running one configuration: the same system-prompt template, or the same tool set when there is no system prompt. It is an observed grouping, not necessarily an agent registered in the console's Agents page.

Review Your agents by spend, What a run costs, Steps against cost, and How runs end to find expensive or incomplete runs. The context panels show what fills prompts, cache use by step, and where long context costs most. Tool panels compare offered tools with called tools and show timing, definition changes and what happens after tool results.

Follow a configuration or run into its details before deciding whether to change a model, prompt or tool set. Check the displayed data-through time when assessing freshness. Options shown as disabled, such as Coding agents when unavailable, cannot be opened.

Performance & errors​

Use this tab to separate provider time, gateway time and request failures.

Where the time goes compares total latency, time waiting on the provider and time added by Quilr. These measurements can overlap, so the parts do not add up to the total.

Review failure causes and affected apps, then inspect streaming endings, rate limiting, output limits and reasoning tokens. Model versions served, Requests by provider, and Same model on different deployments help distinguish a model-wide issue from one provider connection. Select an app or failure entry for the requests behind it.

Security​

Use this tab to review what was scanned, what was found, and how findings were handled.

Review Results, Guardrail actions, What was found, and Coverage by app together. A monitored finding is different from a blocked or redacted request. Found in categories you haven't turned on helps identify findings outside your enabled protection, while Where apps are used from shows the observed source networks.

If Scan completeness isn't available yet appears, completeness is unavailable for that view. It does not mean that nothing was scanned or that every request was protected. See Security guardrails to change the applicable controls.

Attachments​

Review files sent with model requests and the scans that ran on them.

The summary separates requests with files, file counts, attachment bytes, blocked or stripped requests, scanned requests and OCR runs. Use File kinds, Detected file types, Scan outcomes, Data found in attachment text, Image labels, and Why attachments were skipped to investigate coverage.

Select a chart or application to review related requests. Check skipped and incomplete scans as well as detections; a successful model response alone does not show that an attachment was scanned. See Attachment scanning for configuration and availability.

Users​

Find who uses the gateway, which applications identify their callers, and where sensitive data appears.

Switch between People and Smart groups, search for a person, and compare Requests, Est. cost, Sensitive data, or Redacted or blocked. Coverage by app helps you assess identified usage. Sensitive data by person separates prompts from replies, and People → agents → tools connects callers with observed configurations and tools.

Identity-based views depend on the caller information supplied with requests. See Identity and network trust and Agent monitoring to improve correlation.

Read the figures in context​

Costs are estimates, not a provider invoice. Agent/configuration groupings describe observed traffic, and a signal is a starting point for investigation. Read the period, scope, freshness and any unavailable-data message alongside a number.

Overall analytics opens Insights. To inspect individual calls directly, return to the application card and select Logs for Activity > Requests. Interactions groups calls into conversations and Findings shows guardrail detections. App settings remain under the card's Configure actions.

On this page