Skip to main content

MCP usage in Analytics

The MCP Usage tab on the Analytics page shows how your account is used through MCP: by agents your account registered, and by the AI clients people connected, such as Claude, ChatGPT or Cursor. It counts requests, tool calls, refusals and limits over a period you choose. It is read-only: nothing on it changes a setting, a limit or a connection.

Who can see it​

You see the tab when you hold the Edit or Admin permission level on the account itself. Edit or Admin on an organization or group is not enough, because the tab covers the whole account. People with View access do not see the tab at all.

Editors and Admins see the same tab, with one difference, in the People section:

  • Editors see how people use AI clients as totals: per AI client and per scope tier, the number of people connected and active, their requests, tool calls, writes and refusals. Editors do not see any one colleague's use.
  • Admins also see use per person: each person's requests, tool calls by tier, refusals, the AI clients they used (with the requests and the tiers of the calls made from each) and when they were last active.

Nobody sees what was asked or answered. The tab holds counts only: no tool arguments, no results and no timeline of individual calls.

It covers the whole account​

Every other Analytics tab follows the organization or group you are working in. This one does not. MCP use belongs to the account, so the tab always shows the whole account, whichever organization or group you have selected, and says so at the top.

Choosing a period​

The tab has its own period control, separate from the date range of the other tabs. Choose the last 24 hours (shown hour by hour), the last 7 or 30 days (day by day), or the last 13 months (month by month). Periods start and end on UTC boundaries, and every time on the tab is shown in UTC. Your choice stays in the page address, so the browser's Back button and a shared link bring you back to the same period.

Measurement started recently, so a period can reach back further than the data does. When that happens, the tab tells you the day history starts from. If nothing reached the account in the period, the tab says so, and still shows who is connected and your agents' token requests. Earlier days are empty because nothing was measured then, not because nothing happened.

How current the figures are​

Usage is recorded in batches about every five minutes, so the figures can be up to about five minutes behind. The header shows the time of the latest batch. If no batch has arrived for a while, the header reads "data freshness unknown" instead of a time. Use Refresh to load the latest figures. A figure that could not be loaded reads "unavailable" rather than zero.

The Agents section​

This section covers registered agents, which act under their own identity.

Beta. Agent limits, the account's request budget and its concurrent agent capacity, are in beta. Floxar may modify these limits at any time.

  • Requests: the agents' requests through MCP and through the REST API, over the period.
  • Peak demand per calendar minute: the most agent requests in any one calendar minute (UTC). This is demand, not what a limit refused. A rate limit counts a sliding window, so a calendar minute above the request budget does not mean requests were refused.
  • Limits in force: the account's request budget (requests a minute, shared by all of its agents) and its concurrent agent capacity, from its plan or an override set for the account, as of the latest hour. A limit that "changed during the period" is marked. An account admin also sees the account's current limits in the Agent capacity panel on the Agents page. Hours before the account budget took effect on 2026-10-03, when each agent was limited on its own by a per-agent rate tier, read "per-agent tier (retired)"; for a period that ends before then, a "Per-agent rate tiers (retired)" table shows those older limits.
  • Peak active agent trails: the most agent trails active at once, observed when a trail was admitted, against the concurrent capacity in force at the time. A period with no admission is shown as unobserved, never as zero.
  • Capacity outcomes: agent trails refused, queued or moved back to the queue because the concurrent capacity was reached, by where they were started. These are separate from rate-limit refusals.
  • Limit changes: the hours or days in which the account's limits changed, so you can read each peak against the limits in force when it happened. Values are shown in order, the earlier value first with an arrow to the later one, only where the hours they were seen in make the order certain; otherwise they are listed as possibilities, joined by "or". A change from one hour or day to the next is only shown when each had a single value.
  • Agents: each agent's requests, tool calls by scope tier, errors, refusals and when it was last active. Agents no longer have a rate tier of their own; they share the account's request budget. The rate tier column reads "per-agent tier (retired)", or, for a period that ends before the tiers were retired, the tier the agent had then; point at it to see the tiers it had during the period.
  • Tools, agent software and protocol: which tools agents call and how often they fail, the software agents identify themselves with, and the MCP protocol version they use. If an agent still uses the older protocol, a notice says so.
  • Agent token requests: tokens issued to your agents and attempts refused. Only attempts Floxar could attribute to your account are counted.

The People section​

This section covers people working through an AI client they connected.

  • Adoption by AI client: per client, the people with a live connection now against the people who used it in the period. Someone connected who never appears as active may hold a connection worth revoking.
  • Use by AI client: requests, tool calls by scope tier and refusals per client.
  • Writes on people's behalf: tool calls that run or author work, per client.
  • Refusals by AI client: requests refused for scope, connection, rate or permission. Refusals often point at a scope ceiling or client policy worth checking.
  • Use per person: for Admins only, as described above.

Reading the figures​

  • "at least": at very high volume, some activity cannot be attributed to a single agent, person or client. Figures affected by this read "at least" followed by the number. When none of a period's activity could be attributed, a table says how much there was instead of saying there was none.
  • Agents and people are never added together. Each section counts only its own side. A person may also use several AI clients, so per-client counts of people do not add up to the account's total.
  • "Some peaks may read low (cache evictions during these hours)": in hours marked this way, some short-term counters were lost, so the peak demand recorded for those hours can be lower than it really was.
  • "unlimited": shown where a limit has no ceiling.
  • "changed during the hour from …": when a limit changed within the latest hour, the tab shows every value seen in that hour rather than picking one.

Where to act on what you see​

The tab only reports. Account Admins manage agents in the Agent Registry and people's connections under MCP Connections in the account settings, and the tab links Admins to both. To set up a connection or an agent, see Connecting an AI client and Registering an agent.

Last reviewed: 2026-10-03