Skip to content

Observability, rebuilt for Agents.

A full-fidelity observability data lake built for machine-scale investigation, running entirely in your own cloud.

Both SaaS and in-your-cloud deployments supported.

Founded by ex-Netflix engineersBacked by CRV
User asks Claude
Why did checkout latency increase?
4
queries
fired in parallel
0
dimensions
5m
history compared
184 GB
telemetry examined
4 queries. One context window.
— Evidence synthesis —
184 GB
telemetry examinedClaude's context
Trusted by teams at
Aviatrix logoRidecell logoMoxxi logoReadyset logoOlto logoFleak logoBMO logo

A runtime built for machine-scale investigation.

OpenTelemetry writes directly to your object storage. Every investigation gets dedicated compute that runs against your own telemetry. Your agents send questions. They receive evidence. Your data never leaves your cloud.

YOUR CLOUDOpenTelemetry CollectorLogs · Metrics · TracesYOUR OBJECT STORAGEYears of logs · metrics · tracesCardinal RuntimeSpins up for every investigationINVESTIGATE →← EVIDENCEClaudeCodexGeminiCursor

Deploy Cardinal in minutes.

Install the Cardinal plugin in your coding agent, connect, and let it stand up your observability data lake inside your own cloud.

cardinal — claude code
$

Let your agents investigate without compromise.

Traditional observability platforms sample your traces, cap cardinality, age out logs, and throttle queries. These limits keep the vendor’s costs down at human query volumes. Agents investigate at machine scale and hit every one. Cardinal has none of them. Your agents can keep going until the evidence, not the platform, becomes the limit.

Traditional observability
live trace
Why did checkout latency increase?
00:01Querying checkout p99 by region
00:02!Trace sample: 88% dropped by collector
00:03×429 Too Many Requests
00:03Retrying in 2s ↺
00:05Comparing three-month baseline
00:08!No data returned — beyond 14d retention
00:10Segmenting by customer_id
00:13!customer_id dropped — cardinality cap
Investigation degraded
Queries completed
3
Dimensions explored
2
History compared
14d cap
Rate-limited
1
Elapsed
00:19
Outcome
The failure isn’t accidental — it’s architectural.
Sampling, retention limits, cardinality caps, query throttling — these walls make the vendor’s cost predictable. They make your investigation impossible.
Cardinal
live trace
Why did checkout latency increase?
00:01
Comparing latency across regions
00:01
Comparing current vs healthy deployments
00:02
Segmenting by customer_id
00:02
Inspecting retry behavior
00:02
Correlating Kubernetes events
00:03
Comparing three-month baseline
00:03
Searching similar incidents
00:04
Synthesizing evidence
Investigation complete
Queries completed
3,442
Dimensions explored
36
History compared
3 months
Rate-limited
0
Elapsed
00:04
Root cause
payments-api:v84 introduced a retry storm after deployment.
Dedicated compute lets every investigation run at machine scale.

Your telemetry is infinite. Your context window isn’t.

Every production question is a needle-in-a-haystack problem: billions of events, one answer. No LLM can hold the haystack.

Cardinal semantically compresses it. Your agent reasons over the shape of production and pulls raw evidence only where the needle is.

184 GB. 42 log lines. Same answer.

The haystack, semantically compressed
Percentile sketchesPrincipal-component plotsLog fingerprint clustersTrace-flow attributions
Raw
12M points
millions of measurements
1,432 series
high-cardinality series
184 GB
hundreds of GB of logs
18M spans
millions of trace spans
Semantic
compression
Compressed
p50
240ms
p90
480ms
p95
620ms
p99
780ms
max
1.2s
PC1PC2
HTTP %s %s → %d×2,143
db.query timeout after %s×487
retry attempt=%d id=%s×312
cache miss key=%s×128
service · payments-apiRED
webcheckoutcartledgernotifierauth
≈ 40 bytes
5 percentile buckets
≈ 80 bytes
2-axis projection
8 fingerprints
fingerprint clusters
3 sankeys
one each for R · E · D, per service

Compile agentic investigations into mechanical runbooks.

Every agent investigation captures hard-earned operational knowledge. Cardinal compiles successful agent investigations into lightweight runbooks that watch your infrastructure forever — or become building blocks for every future agent.

No exploration at runtime
The search happens once. Every run after is deterministic.

Discovery is expensive. Execution shouldn't be. Once a runbook is compiled, every subsequent run costs a fraction of the original — predictable, bounded, no runaway loops.

Compounding operational memory
Never solve the same problem twice.

Every answered question becomes reusable software. Future agents inherit every investigation your team has already run.

CLAUDE> How is my revenue doing? querying BigQuery · orders_todaytoday $842K −18% vs 7-day avg user-facing SLO statuscheckout-svc p99 940ms breachpayments-svc err 8.4% breach recent commits in those reposcheckout-svc a3f9c2 · pricing-v2 rolloutroot cause: pricing-v2 rejects valid carts> Great Find!> /mechanize Converting this investigation into a runbook…RUNBOOK LIBRARYREADY TO RUNSLO-BREACH-SCANDEPLOY-CANARY-CHECKCOST-ANOMALY-WATCHCACHE-STAMPEDE-GUARDPAYMENT-RETRY-STORMQUEUE-BACKLOG-ALERTREVENUE-DIP-ANALYSIS
Pricing

Transparent pricing. No sales call required.

Same product either way. Pick where Cardinal runs, see the price on the page.

  • Free: Start free.10 GB/day, forever. Every feature. No credit card.
  • Fixed price: Your bill stays predictable.The bill is the same whether you run one query or a million. Retention and cardinality don't move it either.
  • Deployment: Your cloud or ours.Choose the deployment model, not a different product.
  • Support: Buy expertise, not more telemetry.Higher tiers add faster response, architecture guidance, and enterprise support.

Give your agents the evidence they deserve.

Stop building agents on infrastructure designed for dashboards. Give them complete evidence, machine-scale investigations and infrastructure you control.

Compliance
SOC 2 Type IICompliant

Independently audited annually. Your data never leaves your cloud.