How we count AI agent traffic
The definitions behind the figures on /ai, and the live sources they come from. Every source below is public and needs no key.
Live reading
Read from the sources below when this page loaded. A figure that could not be read is left blank.
| Distinct external agents, last 7 days | |
|---|---|
| The same, excluding the single largest caller | |
| Share of 7-day tool calls from the largest caller | |
| Real external tool calls, last 7 days | |
| Recognised AI platforms with a tool call, last 30 days |
Platforms that called MCP tools in the last 30 days
| Platform | Calls | Active days | Last call | Named on /ai |
|---|---|---|---|---|
| Reading the live window | ||||
Agents
An agent is a distinct public client address that made a real external MCP tool call in the trailing 7 days. Crawlers, uptime probes, QA harnesses and DC Hub's own traffic are excluded.
Identity is derived from the client IP, so it is a proxy. Several agents behind one network address count once, and one agent on rotating addresses counts several times.
When one caller makes a large share of the week's calls, we publish the count without it beside the headline figure. The table above shows both.
Platforms
A platform is a recognised AI assistant or tool vendor with at least one MCP tool call in the trailing 30 days. Recognition is an allowlist, so a client we have not classified is not listed even when it calls. A vendor that calls under several client names (for example claude, claude-code and the Anthropic API) is listed once.
The count on /ai is the canonical platform figure. The sentence names a platform only when it called on more than one day and made at least 10 calls in the window, so it says "including" when fewer platforms clear that bar than the count. Every platform in the window is listed in the table above with the same three facts: calls, active days and last call.
Tool calls and crawler requests
These are two different channels. An MCP tool call is an agent asking DC Hub a question. A crawler or citation request is an AI platform reading a DC Hub page, attributed by its User-Agent. We publish them separately and never add them together.
The Gemini series excludes Googlebot (Google Search) from 2026-09-05 and counts Google's user-triggered Gemini fetchers (Google-Agent, Google-NotebookLM, Google-GeminiNotebook) from 2026-09-26. Figures across those dates are not comparable.
Total requests served
The headline on /ai is the all-time count of requests that AI platforms and MCP clients have made to DC Hub: requests from recognised AI platforms, attributed by User-Agent, plus requests to the MCP transport. They are requests, not tool calls, and they are not the agent or tool-call figures above.
It leaves out DC Hub's own internal services, direct API traffic and probe or scanner traffic, and the line under the number prints that excluded count, so the two always add up to everything the server received. The source is /api/public/mcp-count: its total, plus the MCP-transport buckets listed under excluded_buckets.
Platform census
The platform cards in the Live Dashboard tab on /ai list every client that has sent DC Hub any request, all-time: AI assistants, MCP registries, developer tooling and harvesters alike. That is why the census is wider than the platform figure, which counts recognised AI platforms with a tool call in the last 30 days.
A card marked MCP active made at least one tool call in the last 30 days, whoever the client is, so the census can carry more of them than the platform figure counts. Both cover the last 30 days and both need one call, but they do not count the same population. The platform figure counts only clients we recognise as AI platforms, and only calls that came from outside DC Hub, with our own declared test traffic removed. The census counts every client that called a tool, so a registry, or a client we have not classified, shows on its card with its calls and is not added to the platform figure, and a recognised platform shows its calls before that removal.
Probe, scanner, QA and DC Hub's own client names are left out of the census. In the 7-day view of the reach chart, requests to endpoints only our own dashboards poll are removed when the data carries that figure, and the chart says so when it does not.
Outreach
The Outreach tab on /ai reads the ledgers real outbound work writes: partner emails sent, MCP directories that list DC Hub (from the directory listing checks), and agents that registered themselves through the /ai-agents form, the submit_platform tool or the API.
A lane reads Idle when its ledger has no recent entry. Nothing on the tab is a scheduled or estimated count.
Facilities
Facility figures count distinct buildings after duplicate listings are merged, published as a floor. We do not publish a headline facility count until a corroborated count is ready.
Sources
- /api/v1/ai/reach
7-day agents and calls, the largest-caller share, and the basis for each figure - /api/v1/stats/live-proof
The canonical agent and platform figures (thefiguresblock), and tool calls per platform over 30 days with what was excluded and why - /api/agent/stats
Platform totals with their query - /api/public/mcp-count
Requests served to AI platforms and MCP clients, all-time, with the traffic the headline leaves out - /api/ai/tracking
Crawler and citation requests, the per-platform census and the recent request feed - /api/ai-ecosystem/status
The outreach ledgers - /api/v1/stats/canonical
Facility counting rules - /mcp/directories
Every MCP registry and directory that lists DC Hub, with its status