Cloud/AI Governance/Agent Activity
Agent Activity
Every AI client that connects to your fleet, on one screen: who holds which token, who makes the calls, who is slow and who is failing, with versions and transports identified.
Mission Control shows the fleet. Agent Activity shows the actors: every AI client that connects to Vinkius, identified by name, version and transport, with its connections, tool calls, data transfer, latency and errors side by side. Agents are software you did not write, running with your credentials; this screen is the guest list. The console puts its own promise in one line: "Stop guessing which agents matter. See who drives value, who drains resources, and what your connection patterns are really telling you."

Agent Activity
Stop guessing which agents matter. See who drives value, who drains resources, and what your connection patterns are really telling you.
Token Distribution
Connections per client, stacked by tokenTool Usage by Client
Total tool calls · color = latency health| AI Client | Connections | Tool Calls | Avg Latency | Errors |
|---|---|---|---|---|
| Cursorv1.2.3streamable-http | 24 | 912 | 96ms | 4 |
| Windsurfv1.0.2sse | 14 | 327 | 182ms | 2 |
| Claude Desktopv1.2.1streamable-http | 21 | 642 | 124ms | 9 |
| VS Code Copilotv1.1.0sse | 17 | 461 | 268ms | 11 |
| Clinev1.2.4streamable-http | 11 | 272 | 88ms | 0 |
| Zedv1.0.3sse | 7 | 233 | 143ms | 1 |
The mockup runs on the console's Sample Data (the same simulated set the free tier explores, with well known clients such as Cursor, Claude Desktop and VS Code Copilot). On a paid plan the same screen shows your live clients, per organization.
The four counters
- Connections. Active sessions in the period: how many distinct client connections your fleet is serving right now.
- Tool Calls. Total executions, with the derived calls per session right under the number, so a spike in calls with a flat session count reads instantly as "the same agents doing much more".
- Data Transfer. Total payload volume crossing the proxy. This is the number your FinOps conversations start from.
- Avg Session. Average connection duration. Long sessions mean agents working in long-running loops; short ones mean quick questions.
Two charts: who holds which key, and who does the work
Token Distribution stacks each client's connections by the Connection Token they used. This is a security reading, not just a pretty chart: every bar segment tells you which credential each client arrived with. A token appearing under an unexpected client is exactly the kind of thing you catch here, one glance earlier than anywhere else.
Tool Usage by Client ranks clients by total tool calls, and the bar color is latency health: green for fast clients, amber when a client's average latency crosses the 200ms line. The heavy client and the slow client are rarely the same one; this chart makes that visible.
The client table: identified, not anonymous
Every row is one AI client, and the console refuses to treat them as anonymous traffic:
- AI Client with a version chip (v1.2.3 style) and a transport chip (streamable-http or sse), so you know not just which client but which build and which MCP transport it used.
- Connections and Tool Calls per client, in monospace numbers that line up for comparison.
- Data Transfer per client, for the conversation about who consumes the fleet.
- Avg Latency color coded: green under 200ms, amber under 800ms, red after.
- Errors in red the moment they exist, zero in quiet gray when the client is clean.
- Tools and Servers: how many distinct tools the client reached and across how many Connectors, a compact map of how wide its reach is.
- Avg Session duration per client.
Clicking a row opens that client's detail. The empty state is the console's own: "No AI client connections recorded in this period."
From the overview to the actors
The Details link on Mission Control's charts drops you here, and every number on this screen answers a question the overview raises: which client is behind that traffic, that latency, that error count. When something looks wrong here, the next stops in the sidebar are Access Tokens (who holds which key) and Request Failures (what exactly failed).