Argus vs OpenTelemetry for Claude Cowork
If you're trying to monitor a Claude Cowork rollout, the two real options today are OpenTelemetry (Anthropic's official telemetry path) and a capture plugin like Argus. They aren't substitutes — they answer different questions. This page is the honest read of where each one earns its keep.
TL;DR
- OpenTelemetry: cost per user, latency per turn, tool-call counts, routed to your existing observability stack. Aggregate metrics. Not a compliance-grade audit trail (Anthropic's own docs).
- Argus: the actual session — prompts, MCP calls, assistant outputs, version hashes — with replay, annotation, and QA scoring per turn. Per- workspace storage; designed to fill the audit-trail gap Cowork itself doesn't cover.
You want both. OTel for the quantities, Argus for the qualities.
Side-by-side
| Question | OpenTelemetry | Argus |
|---|---|---|
| How much did Cowork cost my org last month? | ✓ | partial (via OTel passthrough) |
| Which users are heavy / light? | ✓ | ✓ |
| What did this specific session do, turn by turn? | ✗ | ✓ |
| Which skills fired, in which versions? | ✗ | ✓ |
| Did the user accept the answer on the first turn? | ✗ | ✓ |
| Which MCPs are installed across the team? | ✗ | ✓ |
Can I tag a session as wrong and re-find it later? | ✗ | ✓ |
| Compliance-grade audit trail? | ✗ (per Anthropic) | ✓ (with redaction) |
| Privacy posture | Aggregated metrics only | Per-workspace; opt-out per session |
| Setup time | OTel collector + ~4 env vars | Plugin install + sign-in |
When OpenTelemetry is enough
You're a small team running Cowork casually, you care mostly about spend control, you have an observability stack already (Grafana, Datadog, SigNoz). Wire the four OTel env vars and route the events. You'll see cost-per-user and have a basis for budgeting.
When you need session capture
You're shipping Cowork into someone else's organization, you have a QA loop to run, you have skills you want to refine over time, or you have a compliance team that will ask "what did this agent do on this date." OTel will tell you that it happened — you need a capture layer to tell you what.
Can I use both?
Yes — and you should. Argus consumes OTel events on the same ingestion endpoint, so the cost and latency data flows into the same dashboard as the session content. The two are merged per session, side by side, in the replay view.
What we don't do
Argus is a Cowork-shaped tool. We're not a general LLM observability platform — if you want Anthropic + OpenAI + Gemini all in one trace view, Langfuse or LangSmith is the better fit. We focus narrowly on the Cowork / Claude Code session surface where the existing tools are weakest.