collector nominal · 240k spans/s
COLLECTOR · NOMINAL · 99.97%|12M spans/min

Watch everyprompt, everytoken, every retry.

Argus is the observability layer for LLM-shaped workloads. Trace each agent step, score each response, replay each span. Find the bad agent before your customer does.

Install collectorRead SDK
8.3
min MTTR
94.6%
caught
12
SDKs
240k
spans/s
trace · 7e08af2c · live
↻ 0.8s
#
span
model
ms
tok
01
intent.classify
haiku-4
84
218
02
retrieve.context
vec-search
47
—
03
tool.lookup_account
—
213
—
04
model.call
sonnet-4.6
612
1840
05
tool.refund_check
—
318
—
06
tool.refund_check (retry)
—
184
—
07
model.call
sonnet-4.6
487
1247
08
grade.quality
rubric-v3
47
84
2.0s
total
$0.0182
cost
0.91
quality
Mercury OS●Substrate●Northwind Labs●Foundry 47●Marlow Workwear●Plover Studio●Verge Editions●Field Notes Co.●Linnea Bakery●Helvetia Cartographic●Mercury OS●Substrate●Northwind Labs●Foundry 47●Marlow Workwear●Plover Studio●Verge Editions●Field Notes Co.●Linnea Bakery●Helvetia Cartographic●
02 / Surveillance surface

Four panels. Twenty-four eyes.

SPAN TREE
›agent.run2014ms
›intent.classify84ms
›retrieve.context47ms
›model.call612ms
›tool.refund_check318ms
›tool.refund_check.retry184ms
›grade.quality47ms

A queryable graph of every prompt, tool call, retry, and grading step. Filter to a single customer in milliseconds.

QUALITY DRIFT
drift
ALERTS
WARN · 04:18
refund_check p99 → 1.4s
OK · 03:47
retry rate normalised
OK · 02:04
collector failover ok
COST SURVEILLANCE
Mercury OS
$8,427
Substrate
$6,184
Northwind Labs
$4,218
Foundry 47
$2,104
Linnea Bakery
$412
REPLAY
replay span
tool.refund_check / 7e08af2c
last replay · 213ms · cost $0.0021
03 / In production

Three teams. One console.

Incident response, quality drift, cost surveillance. Three jobs that used to live in three tools. Now one timeline.

01
Incident

From customer ticket to root cause in eight minutes.

Argus reconstructs the whole agent run — every prompt, every tool call, every retry — and surfaces the divergent step. The 3am debug stops being a tradition.

0
minute median MTTR
02
Quality drift

Catch a regression before customer support does.

Score every production response against your rubric. When the line moves, Argus pages the right team and attaches the offending sample. No spreadsheet required.

0
% of regressions caught pre-customer
03
Cost surveillance

Find the prompt that quietly tripled your bill.

Per-prompt, per-model, per-customer cost roll-ups. The expensive outlier rises to the top of the list, with the offending input attached.

0
% reduction in token bill p99
04 / Vs APM

APM treats LLMs as a black box. We don't.

Capability
Argus
Generic APM
Trace surface
Per-token timeline
Aggregate counters
Quality scoring
Built-in rubric runner
Bring your own
Replay
Replay any span, any model
Read-only logs
PII handling
Field-level redaction
Disable logging
Alerting
Per-customer thresholds
One-size global
field reports

“Argus turned an opaque agent into a debuggable program. We close incidents in single-digit minutes now.”

Kohei Tanaka
Site Reliability, Mercury OS
0k
spans / second
0m
median MTTR
0%
regressions caught
0%
collector uptime
05 / Pricing

Spans are cheap. Outages are not.

Recon
$0free forever

For solo developers and prototypes.

  • 100k spans / mo
  • 7-day retention
  • Single project
  • Community Discord
Start tracing
Watch
$112per seat / month

For teams with production agents.

  • 12M spans / mo
  • 90-day retention
  • Quality rubric runner
  • Replay any span
  • Slack + Linear alerts
  • Business-hour support
Start 14-day trial
Sentinel
—annual contract

For regulated and high-volume teams.

  • Unlimited spans
  • On-prem collector
  • Field-level PII policy
  • SAML / SCIM
  • Named SRE
  • 99.97% SLA
Talk to sales
06 / FAQ

Honest answers to honest questions.

APM treats an LLM call as a single span. Argus treats it as a graph: prompt, tool calls, model outputs, retries, and grading — each as a queryable span. You can ask 'show me all spans where the model picked tool refund_check after retrieving a contract clause' and get an answer in seconds.

Recent ops

all releases
Apr 25
alerts
Per-customer thresholds for cost and latency, with sample attached to the page.
Apr 16
sdk
Python SDK 5.2 ships zero-config auto-instrumentation for LangGraph.
Apr 04
trace
Span tree gains tool-call grouping and retry collapse for readability.
Mar 22
regions
Frankfurt collector is generally available; eu-central-1 ready for traffic.

Trace it before it pages you.

One npm install. One env var. Hundred thousand spans on the house. No card.

Install free Talk to SRE