Detection layer for LLM pipelines

Your agent
is failing.
Silently.

Cernova learns what each step normally does and flags the moment one breaks — the failures your logs record as success.

Start free →See how it works
01 · Detect

It learns as
your agent runs.

Every step gets a baseline from its own history — no thresholds to configure. Cernova knows what normal looks like for this exact call.

profile #d4b97b847 clean samples
02 · Diagnose

Catches what your
logs call success.

When a call breaks the pattern you get the evidence — what Cernova observed, and what it expected.

observed 3400msexpected ≤ 240
03 · Locate

Straight to the
exact line.

Every call records where in your code it ran — from “this step broke” to the file and line, for you or an agent in your editor.

anomaly generate-replyat workflow.ts:256
04 · Connect

Alerts where you
already work.

Pull anomalies into Claude Code over MCP, or push signed events to Slack, Sentry, and your own webhook. Keep your stack.

@cernova/mcpslacksentry
// the dashboard

Every step, learned
and watched.

Cernova splits your agent into steps and learns two things about each — its normal behavior and its expected output shape — then flags the one that drifts. This is the real console.

cernova · support-agent · steps
3 steps monitored · 1 still learning30 calls per step to start
generate-replycritical198 calls▲
behavior
latency890ms→3120ms3.5× normal
output shape
Free-form text output — no structural keys.
retrieve-contextdegrading156 calls▼
classify-intenthealthyoutput shape to review214 calls▲
output shape
learned from 42 outputs · json · valid JSON 98% · watching — confirm to enforce
keytyperequiredconstraints
categorystringrequired∈ general · billing · technical · account
confidencenumberrequired0 – 1
summarize-ticketlearning12 calls▼
// claude code · mcp

Ask your agent
what broke.

The MCP server puts Cernova inside Claude Code. Ask in plain language; it pulls the run, the evidence, and the call site — then tells you the fix.

claude code — support-agent
› which step is failing in my support agent?
● I'll check Cernova for anomalies.
⏺ cernova · list_anomalies (MCP)
⎿ 1 run · critical · score 600
generate-reply · latency_iqr_fence
⏺ cernova · get_run (MCP)
⎿ run d33f575d · 3 calls
observed 3400ms · expected ≤ 240
code_filepath: sample-app/workflow.ts:256
● The failing step is generate-reply. Latency spiked to 3400ms against a 240ms baseline — a 14× regression — at workflow.ts:256. The retriever is returning oversized context; scope it to the intent before the reply call.
› ▋
// keep your stack, add detection
Traces in
Anthropic SDKOpenAI SDKLangChainOpenTelemetryVercel AI SDKManual ingest
Alerts out
SlackSentrySigned webhooks
Data out
Read APIMCP · Claude CodeLangfuse / LangSmith import

Green dashboards.
Broken agents.

See what your logs record as success. Start charting the dark.

Start free →Read the docs
Cernova · the detection layer for LLM pipelines · set in Space Grotesk & IBM Plex Mono