Now in private preview

Stop running AI pipelines
blind.

LLM observability for AI pipelines — quality, cost and stability control, built for the teams shipping LLMs in production.

Insights / Issuesworkspace: support-agentSample data7d

Issues

Occurrences that share a fingerprint are grouped into one Issue.
Open issues
12
Occurrences
52,480
Attributed cost
≥ $312.40
20,210 of 52,480 priced
IssueTrendOccurrencesTracesImpact
CostToolRepeatsupport-agent/act_<num>New
act_2 repeats search_orders 6×
TOOL-001 · node · in range: last 12m ago · first 5d ago
18,636
6,212
$186.40
18,636 of 18,636 priced
CostOversizedResponsesupport-agent/act_<num>New
search_knowledge_base returns 422544 chars into execute_tool search_knowledge_base
TOOL-004 · node · in range: last 41m ago · first 6d ago
318
318
$94.20
318 of 318 priced
ReliabilityToolFailuresupport-agent/act_<num>New
execute_tool search_orders: search_orders fails with Timeout
ERR-001 · node · in range: last 8m ago · first 5d ago
1,418
860
+11h 49m
Cross-domainVersionRegressionsupport-agent/New
invoke_agent support-agent: 1.5.0 regresses cost vs 1.4.0
DEP-001 · node · in range: last 4m ago · first 5d ago
29,540
29,540
unknown

The LLM & AI agent observability platform

Quality. Cost. Stability.

One platform for AI observability — every signal your AI pipeline emits, from the first token to the last billing line.

Quality

Trust every answer your model ships.

Catch regressions before users do — with online LLM evaluation and human feedback wired into every turn.

  • LLM evaluation: faithfulness, relevance, context recall
  • Hallucination & PII leak detection (regex + LLM)
  • Human 1–5 feedback loops with sampling controls
  • Per-conversation eval pass-rate with drill-down
Cost

Know what every token costs.

Rollups by model, provider, user and prompt. Projected EOM. Budgets and alerts before you bleed.

  • Spend MTD, EOM, cost per conversation
  • Daily spend stacked by model / workspace / user
  • Token cache hit-rate and savings
  • Budgets with multi-channel alerting
Stability

Stay on top of every incident.

Traces, errors, SLOs and on-call paging — the SRE stack your AI agents have been missing.

  • Distributed traces across agents and tool calls
  • Alerts with PagerDuty / Slack / on-call routing
  • SLOs with burn-rate and 30d error budgets
  • Provider downtime feeds (OpenAI, Anthropic, etc)

A look inside

LLM observability built for the way AI teams debug.

Users find problems first

Issues find problems automatically, without manual trace review. When a user complains, Conversations shows the full conversation, model and tool calls with inputs and outputs, and quality scores

Slow response, unknown cause

The trace shows LLM calls next to SQL, HTTP and queues: you see which step takes the time

LLM Observability / Conversationsworkspace: support-agentSample data

Conversation

sess_7f21
Conversation timeline· 2m 26s wall-clock · 12 operations
Conversation thread
Scores
  • relevance0.6
  • tonepolite
Where is my order 4417?
Assistant model-secondary
Order 4417 shipped on Monday and arrives on Thursday.
Send it to 12 Harbour Street.
Assistant model-primary
Sorry, what is your order number?
I already told you. 4417.
Assistant model-primary
I could not update order 4417. Please contact support.

Pricing

Pay as you go.

Transparent, usage-based. No seats. No platform fee. Bring your storage if you want.

TripleCloud pricing pipelineYour OpenTelemetry exporter pushes spans into Ingest (€0.25 per GB ingested), where the LLM processor parses them (€2.50 per GB processed) before they are written to Store (€0.25 per 10GB per month). Stored spans are then evaluated by Score (€0.75 per 10k span evaluations) — evaluation orchestration only; bring your own judge-model provider.
OpenTelemetry
Ingest€0.25 / GB
Process€2.50 / GB
Store€0.25 / 10GB·month
Score€0.75 / 10k
See full pricing details
€0.25/ GB
Ingest
Raw events, logs and conversation payloads.
€2.50/ GB
Process
Trace ingestion, parsing and enrichment across agents. Non-LLM spans will not be processed.
€0.25/ 10GB·month
Store
Hot, queryable retention. BYOStorage available.
€0.75/ 10k scores
Score
Evaluate LLM spans for their result. Orchestration only - BYO model provider.
SSO · Dedicated instance · Certifications available

Open source

The engine is open source.

TripleCloud runs on IceGate — an Apache-2.0 observability data lake engine. Self-host the exact engine we run, or let us operate it. No black box, no lock-in.

OpenTelemetry-native

Point any OTel SDK at IceGate. Exactly-once delivery, straight into open Parquet — no agents to rewrite.

Open formats, no lock-in

Apache Iceberg, Arrow and Parquet under a Rust-native DataFusion query engine. Your data stays portable.

Your storage, your bucket

WAL, catalog and data all live in S3-compatible object storage. Self-host the whole thing under Apache 2.0.

Star on GitHub

Pricing comparison

Trace the whole system. Pay the AI premium only for AI.

Adjust your span mix and scoring volume to compare the estimated monthly cost with Langfuse and Braintrust.

Regular ingest only €0.25/GBNo seats · No platform fee
3 / 10
0246810
1
0246810
Cohort
Total data
Regular spans
AI spans
Traces
Scores
TripleCloud
Langfuse Core
Langfuse Pro
Braintrust Starter
Braintrust Pro
Small
3.7 GB
0.7 GB
3 GB
50k
50k
€12
€69
€239
€111
€249
Medium
37 GB
7 GB
30 GB
500k
500k
€125
€451
€621
€1 369
€1 020
Large
370 GB
70 GB
300 GB
5M
5M
€1 245
€3 931
€4 101
€13 951
€8 769
Small50k traces / mo
Total data3.7 GBRegular spans0.7 GBAI spans3 GBTraces50kScores50k
TripleCloud€12
Langfuse Core€69
Langfuse Pro€239
Braintrust Starter€111
Braintrust Pro€249
Medium500k traces / mo
Total data37 GBRegular spans7 GBAI spans30 GBTraces500kScores500k
TripleCloud€125
Langfuse Core€451
Langfuse Pro€621
Braintrust Starter€1 369
Braintrust Pro€1 020
Large5M traces / mo
Total data370 GBRegular spans70 GBAI spans300 GBTraces5MScores5M
TripleCloud€1 245
Langfuse Core€3 931
Langfuse Pro€4 101
Braintrust Starter€13 951
Braintrust Pro€8 769

Per trace: 7 regular spans + 3 AI / LLM / agent spans + 1 score.

Estimated byte mix: 19% regular · 81% LLM.

Assumptions & method
  • Illustrative size model: each cohort fixes 50k / 500k / 5M traces per month and 10 spans per trace. Data volume is derived from 2 KB per regular span and 20 KB per LLM span; 1 KB = 1,000 bytes and 1 GB = 1,000,000,000 bytes.
  • Scores per trace are independent of the LLM-span count: multiple evaluators may score the same span or trace.
  • TripleCloud totals include 90-day hot retention (Store €0.025/GB·mo, steady-state); BYOStorage is available. Competitors are shown at their plan’s included retention — Langfuse Core 90 days, Langfuse Pro 3 years, Braintrust Starter 14 days, Braintrust Pro 30 days.
  • Langfuse units are traces + observations/spans + scores and use its published graduated tiers. Braintrust is calculated from processed GB + scores; the Pro subscription fee is included.
  • Competitor list prices were reviewed on 10 July 2026 and are in USD; for a side-by-side display only, the table uses €1 ≈ $1. Positive totals below €1 are shown as <€1; all others are rounded to the nearest whole euro after calculation. Taxes, discounts, model/token charges and token credits are excluded.
  • This is an illustrative estimate, not a quote. Workload assumptions, included retention, taxes, discounts and actual billing may differ.

Team

Built by engineers who have solved infrastructure problems at scale.

A small, senior team building IceGate and TripleCloud.

Sergei Prosvirnin

Sergei Prosvirnin

Co-founder

I’ve spent 10+ years building high-load technical products across fintech, ride-hailing, marketplaces, and analytics — from launching new payment systems to building anti-fraud infrastructure at inDrive and large-scale systems handling millions of products. TripleCloud is the observability platform I wanted as an engineering lead: open, practical, infrastructure-friendly, and built for production LLM pipelines.

Evgenii Mineev

Evgenii Mineev

Co-founder

I’ve spent more than 15 years building cloud infrastructure, managed database platforms, monitoring systems, and high-performance backend services at Yandex and Nebius. At TripleCloud, I’m building the kind of LLM observability platform I would want to use myself: open, reliable, infrastructure-friendly, and designed for teams that need to understand what happens inside production AI pipelines.