See what your AI agents actually cost.
OpenTelemetry for Claude Code. Preconfigured dashboards, ingest-time rollups, ready in 5 minutes.
No credit card required. Free for teams under 10 devs. EU residency on every tier.
Native support for
- Claude Code
- Claude Agent SDK
- Claude Cowork
Built for Claude Code teams
The 90% of agent observability you actually look at.
Generic OTel backends ingest fine. They don't know what a subagent is. Metric AI is shaped exactly like Claude Code, Agent SDK, and Cowork.
Cost per dev, per repo, per prompt
Tokens × model price, rolled up by user, repo, and prompt id. Surface the top 1% prompts burning your budget.
Cache hit % per developer
cache_read / cache_creation / input ratios per dev. Coaching has a target. Devs below 60% are burning money.
Subagent tree viewer
Parent/child span chain rendered for any prompt id. See when claude waits for approval, which tools fan out, how deep recursion goes.
cowork support
Approval flows from Cowork
When your team uses Cowork's human-approval gates, every approve/deny event lands in your dashboard, correlated with prompt id and user. The other backends ingest these as opaque log lines; we render them as a flow.
Active developers · DAU / WAU / MAU
Distinct users per day / week / month. Validates seat ROI. Find devs who got onboarded but stopped using it.
Slack & webhook alerts
Daily spend exceeds X. Error rate over Y%. Retry exhaustion spike. Cron evaluates every 15 min, cooldown protects against alert storms.
Setup
Three steps. ~5 minutes. Devs do nothing.
One bearer for the whole team, pushed via Claude for Teams managed settings. No per-machine work.
01 · admin
Generate the org token
Settings panel in the dashboard. One click. The plaintext is
shown once along with a pre-filled managed-settings.json
snippet ready to paste.
02 · admin
Paste into Claude.ai admin console
Claude Code → Managed settings
One paste. Settings fetch on each claude startup.
No MDM, no SSH, no per-dev edits. (No Claude for Teams? MDM
works the same.)
03 · everyone
Open the dashboard
Sign in with your email — six-digit code arrives via Resend. Nine panels, real numbers within 15 minutes. Subagent tree viewer for any prompt id.
vs the alternatives
Generic backends ingest fine. They don't model your agents.
FAQ
Common questions.
Does this work with my non-Claude services?
Metric AI is built for Claude Code (Agent SDK and Cowork inherit the same dashboards). We accept any OTLP traffic but we don't pretend to be a generic backend — for your microservices and databases keep using Datadog or Grafana.
Do you store our prompts and responses?
No. Metric AI never custodies raw prompt or response bodies. Tokens, costs, span durations, prompt IDs — yes. Bodies — no. The Forensics tier signs URLs into a bucket you own (R2 or S3) and renders from there. Compliance burden stays where it belongs.
How does this compare to Anthropic's built-in monitoring?
Anthropic ships OTel instrumentation across Claude Code, Agent SDK, and Cowork — but explicitly punts the backend to "bring your own". Metric AI is that backend, shaped specifically for the agent loop.
EU data residency?
Yes — Workers, D1, and Analytics Engine all run in EU regions on day one. No US round-trip. Available on every tier.
What if Anthropic ships their own dashboard?
They've said publicly they won't. Even if they do, our edge is price + EU residency + cost-attribution dashboards generic backends won't replicate.
Do you support OTLP/protobuf?
Yes — both OTLP/JSON and OTLP/protobuf, all three signals (traces, metrics, logs). Standard collector config works out of the box.
How long to set up a 50-dev team?
Five minutes. Admin generates one org token in the dashboard,
pastes the rendered snippet into the Claude.ai admin console
(Claude for Teams → Claude Code → Managed settings), saves.
Every dev's next claude session picks up the
bearer automatically — no per-machine work. (No Claude for Teams?
Same snippet via MDM works the same.)
See your numbers in 15 minutes.
Free for teams under 10 devs. Drop a file, refresh, done.