Skip to main content
The Metrics Dashboard shows how your organization uses the AI Gateway: who is spending what, which models and MCP servers they call, how agents behave, and what guardrails catch. Open it from AI Gateway → Metrics. The dashboard has five tabs. Overview answers “who” questions about people, teams, virtual accounts, and agents. Models, MCP, Agents, and Guardrails each focus on one kind of resource. Every tab follows the same layout: summary cards at the top, a trend chart, and a searchable table. The time range, filter bar, and Refresh button sit at the top of every tab. What you can see depends on your data access rules. The My Usage and My Teams Usage pinned filters narrow any tab to your own traffic or your teams’ traffic in one click.

​
Overview

The Overview shows total spend, request volume, success rate, requests blocked by rate or budget limits, and guardrail activity across LLM, MCP, and agent traffic. The Usage Directory below the cards lists every user, team, virtual account, and agent that sent traffic in the selected range, sorted by spend. Switch the column set to see LLM cost, LLM performance, MCP, guardrail, or agent metrics for the same rows. The LLM Cost view includes prompt-cache hit rate, cost savings from caching and auto-routing, and how many requests were rate- or budget-limited. Use it for chargebacks, to find your largest consumers, or to check one team’s usage.
Overview tab showing spend, requests, success rate, quota-limited, and guardrail cards above the Usage Directory table of teams and users

Overview with summary cards and the Usage Directory

​
Models

The Models tab covers LLM traffic: spend, requests, tokens (including the cached share), success rate, and latency, with P95 time to first token. The trend chart can show cost, requests, tokens, errors, error rate, or P95 latency, either per model or as a single aggregate line. Toggle between Models and Virtual Models to see usage by the provider model that served the request or by the virtual model that routed it. The models table has cost, performance, and reliability views. A cost-versus-requests chart at the bottom shows which models are both expensive and heavily used.
Models tab showing spend, requests, tokens, success rate, and latency cards, a cost-over-time chart broken down by model, and the models table

Models tab with cost over time and per-model usage

​
MCP

The MCP tab tracks traffic through the MCP Gateway: total MCP calls, tool calls, active servers, success rate, tool-call latency, and the number of callers. The call-volume chart can be stacked by server, tool, MCP method, or caller, or split into successes and failures. The MCP servers & tools table lists each server with its call share, error rate, and P50/P90/P99 latency. Expand a server to see its individual tools, or switch to a flat list of tools to find the slowest or least reliable ones.
MCP tab showing MCP calls, tool calls, servers, success rate, tool-call latency, and callers, with a stacked calls-per-minute chart and a table of MCP servers

MCP tab with calls per server and the servers and tools table

​
Agents

The Agents tab reports agent runs, how many agents were active, success rate, and run-time percentiles. The chart shows runs, errors, or P95 latency per agent. The table lists each agent’s runs alongside the model calls, MCP calls, and subagent calls it made, so you can see what each agent depends on.
Agents tab showing runs, agents, success rate, and run time cards, a runs-over-time chart, and a table of agents with model, MCP, and subagent calls

Agents tab with runs over time and per-agent activity

​
Guardrails

The Guardrails tab shows how many checks your guardrails ran, how many requests they flagged, mutated, or blocked, how much latency they added, and how many users and applications triggered them. The chart shows these outcomes over time, by outcome or by guardrail. The guardrails table lists each guardrail with where it runs, its check count, its flagged, mutated, blocked, and failed counts, and its latency percentiles. A second table, People & apps that trip guardrails, ranks the users and virtual accounts with the most flags or blocks and lists the guardrails they trigger most.
Guardrails tab showing evaluations, flagged, mutated, blocked, added latency, and users/apps triggering cards, an outcomes-over-time chart, and the guardrails table

Guardrails tab with outcomes over time and per-guardrail results

​
Filter and drill down

The search bar at the top of each tab filters every card, chart, and table on that tab. Filter by user email, team, virtual account, conversation ID, or any custom metadata key sent with requests. You can save filters you use often and reuse recent searches.
Filter dropdown listing User Email, Team, Virtual Account, Conversation ID, and Metadata fields alongside saved filters and presets

Filter by user, team, virtual account, conversation, or metadata

Click any row in a table to open its detail view. A team’s detail view, for example, shows the team’s spend, requests, success rate, limits, and guardrail activity, followed by a per-member breakdown.
Detail panel for the engineering team showing spend, requests, success rate, quota-limited, and guardrail cards, cost over time, and usage per team member

Detail view for a team, with usage per team member

​
Export and API access

To download aggregated metrics as CSV, choose Export Data from the ⋮ menu on supported tabs. To fetch the same data programmatically, see the API guides for model, MCP, guardrail, cache, routing, and agent metrics.
Classic view at the top of the dashboard opens the earlier metrics dashboards, which still include the detailed routing (routing rules, rate limits, budget limits) and semantic cache charts.