vitel — Vitals for AI Agents¶

Problem: AI agents ship code and call it done without ever feeling whether the running system is healthy — latency, errors, saturation, error budget. Result: vitel gives them interoception — turn telemetry into a pass / warn / fail verdict, so the agent won't claim done while vitals are failing.

pip install vitel
vitel demo                       # no API key required

vitel demo grades a synthetic degrading service, prints a FAIL report (error budget burning — deterministic, no LLM key), then grades the fixed version and prints PASS.

What it does¶

Capability	What you get
Grade & report	A machine verdict (`pass`/`warn`/`fail`) + issues grounded in a `Metric` (name, window, observed vs threshold, burn-rate) — deterministic, no key needed.
Match an SLO	Grade telemetry against an `SLO` / error budget — `must: error_rate < 1%`, `must: error_budget_remaining > 20%`, `should: p99(latency_ms) < 300ms`.
Analyze & critique	`analyze` adds anomaly detection and an optional LLM critique (local Ollama) — egress only here; `check` stays offline.
Watch over time	`watch` grades behavior over a window — slow latency regression, resource leak, error spike, flatline/liveness.
Vitals → brain handoff	A distilled `{perceived, next_action, todo}` signal any agent/brain acts on.

Where to go next¶

:material-rocket-launch: Quickstart — install, first grade, the demo.
:material-shield-check: Integration — CI deploy-gate + Verel registration.
:material-api: API — check · analyze · watch · render · perceive.

Deterministic by construction¶

check never touches the network — thresholds, SLO conformance and error-budget burn-rate are computed entirely from the series you give it. Egress happens only on analyze (and only to a backend you name), and on the metric backends you opt into. The base install is light (JSON/CSV + thresholds, no heavy deps) — sources live behind extras: [prometheus], [otel], [cloud], [psutil].

Eyes, ears, brain & vitals¶

vitel is the vitals / interoception sense. It pairs with AgentVision (eyes — confirms it renders), Audel (ears — confirms it sounds right), and Verel (the brain — nothing is "done" until a grader returns a verdict). All four speak the shared agentsensory contract — one Verdict, one Report, one Handoff. vitel confirms the running system is actually healthy.