Cloud/AI Governance/Request Failures
Request Failures
The incident log of your fleet: 342 failures out of 12,480 requests in the sample, each one attributed to Agent, Upstream or Vinkius, with the worst server and worst tool named and a health column that audits Vinkius itself.
When something breaks, most dashboards tell you that it broke. Request Failures tells you who: every failed request is attributed to Agent (the caller sent something invalid), Upstream (the MCP provider failed) or Vinkius (the proxy itself), and the screen names the worst offender on both sides. The console's own line: "Trace every failure back to its source. See the exact combination of server and tool behind the noise, and whether things are getting better or worse."
Request Failures
Trace every failure back to its source. See the exact combination of server and tool behind the noise, and whether things are getting better or worse.
Failure Timeline
Errors (red) · Agent / Upstream / Vinkius ratesTransport Distribution
Connection protocol breakdownErrors by Server
Stacked by Agent / Upstream / VinkiusError Distribution by MCP Server
| MCP Server | Requests | Errors | Rate | Health |
|---|---|---|---|---|
| Notion | 1,842 | 96 | 5.21% | Moderate |
| GitHub | 1,684 | 78 | 4.63% | Moderate |
| Slack | 1,204 | 54 | 4.49% | Moderate |
| Jira | 986 | 47 | 4.77% | Moderate |
| Linear | 842 | 31 | 3.68% | Moderate |
| Stripe | 691 | 18 | 2.6% | Moderate |
| HubSpot | 578 | 12 | 2.08% | Moderate |
| Vercel | 412 | 6 | 1.46% | Healthy |
The mockup runs on the console's Sample Data: 342 failures in 12,480 requests, a 2.74% failure rate. On a paid plan the same screen audits your live traffic, per organization.
The four counters
- Total Errors with the request volume behind it, red the moment they exist.
- Failure Rate on a color coded bar (same thresholds everywhere: zero is green, under 5% amber, beyond red) with the three way split right under it: Agent / Upstream / Vinkius. In the sample, 239 failures were the calling agent's fault, 86 the provider's and 17 Vinkius's own.
- Worst Server, the most erroring Connector, named.
- Worst Tool, the most erroring tool, in monospace, ready to search.
Three charts, one story
The Failure Timeline plots errors over the period, so "is it getting better or worse" is a shape, not a debate. The Transport Distribution breaks the traffic down by connection protocol (in the sample: streamable-http, sse and stdio), because a failure pattern that only shows on one transport is a protocol problem, not a provider one. Errors by Server stacks each Connector's failures by the same three causes, so the biggest bar comes pre-diagnosed.
The table that audits the auditor
Error Distribution by MCP Server lists each Connector with its requests, errors and colored rate. Then comes the column that sets Vinkius apart: Health grades the server only by the errors Vinkius itself caused. Your 4xx mistakes and the provider's 5xx outages do not count against the server here; if this column turns amber, it means the proxy in the middle contributed errors, and that is Vinkius's problem to fix, not yours. A dashboard that grades its own operator is a rare thing; this is what "the buck stops here" looks like as a table column.
The empty state is the console's own: "No error data recorded in this period."
The last report, and what closes the loop
This is the final Reports screen: with it, the eight together cover the fleet from every angle: what happened (Mission Control), who asked (Agent Activity), what answered (Server Traffic and Tool Reliability), which credentials (Access Tokens), what it cost (AI Spend) and how protected it was (Security Posture). When this screen shows a failing server, the Settings group is the response: Connector Policy to restrict it, DLP Protection, FinOps Guard and Circuit Breaker to keep it safe, cheap and alive.