Get pinged when your AI calls fail, stall, or start burning money.

Vergilant is a smoke detector for production LLM calls. Sixty seconds to set up. Then it stays out of the way. That’s the point.

The free tier has all three alert types. No card required.

14:03 · the setup

One URL change. That’s the integration.

  • Swap the base URL and add one header.
  • Anthropic and OpenAI, streaming included.
  • Responses come back untouched. Your API key passes through, never stored.
The whole integration
from anthropic import Anthropic client = Anthropic(      base_url="https://api.anthropic.com",+     base_url="https://proxy.vergilant.dev/anthropic",+     default_headers={"X-Monitor-Key": "lm_…"},) # Done. Every call is now watched.

14:05 · watching

Alerting isn’t a feature here. It’s the whole product.

No dashboards to build, no evals to configure. Two tripwires armed at signup, a third whenever you want it.

modellatencytokenscoststatusrequest body

Two armed at signup, one opt-in

  • error_ratearmedPast 10% failures in 15 minutes. The ping says why: “mostly 429 rate_limit.”
  • cost_spikearmed5× your 7-day baseline, learned automatically. No budgets to guess, no thresholds to tune.
  • silenceopt-inFor when traffic just stops. No error log will ever show it.

Meanwhile, in myapp-prod

14:04:12claude-sonnet-5 · 200 · 1.4s$0.01
14:04:51gpt-4o · 200 · 0.8s$0.02
14:05:22claude-sonnet-5 · 200 · 1.1s$0.09
14:05:58claude-sonnet-5 · 200 · 0.9s$0.31

nothing to report

Your agent is stuck in a loop.

Spend is 5.2× your baseline and climbing.

VergilantAPPjust now
Cost spike: myapp-prod
Spend is 5.2× your 15-minute baseline. Looks like a loop.
$1.84 last 15 min · $0.35 typical

Caught four minutes in, with zero config. No budgets to guess: Vergilant learned this project’s baseline on its own.

Sixty seconds to set up. Free tier included. No card required.

14:11:04 · resolved

Total damage: $1.84.

Sixty seconds of setup, paid for itself in four minutes. Back to work.

We store the envelope, never the letter.

There is no code path that writes a request or response body anywhere. The proxy is open source, so you don’t have to take our word for it.

What we see

  • Model name claude-sonnet-5
  • Latency total & time-to-first-token
  • Token counts input / output
  • Estimated cost
  • Status codes 200 · 429 · 500

What we never see

  • Your prompts
  • Model responses
  • Your users’ data

Never stored, never logged. Verifiable in the open-source proxy.

Free to start. A non-decision to keep.

Priced per project, not per seat. One caught loop pays for years.

Free

$0

The full experience, at hobby scale.

  • 1 project
  • 5,000 requests / month
  • All three alert types
  • Discord & email alerts
  • 7-day metadata retention
Start free

Hit the free cap? That usually means real traffic. Upgrading is one click, and nothing about your setup changes.

Back to quiet. That’s the whole product.