Get started
What is ClawMetry?#
You are running coding agents. Possibly several, possibly across several machines, almost certainly across more than one vendor: Claude Code in one terminal, Cursor in the editor, a Codex run on a server, an OpenClaw agent answering Telegram messages at 3am.
Each of those writes a detailed record of what it did — transcripts, tool calls, token counts — into its own private directory, in its own format. None of them shows you the others. None of them tells you what the whole fleet cost last week. And none of them tells you when one of them has been looping on the same failing command for forty minutes.
ClawMetry reads all of them.
The three jobs#
Observation. Find every agent runtime on the machine, read what it wrote, normalise it into one event schema, and store it locally. This is read-only: ClawMetry opens those files read-only and never writes into an agent's directory. See How detection works.
Understanding. Turn that stream into things you can act on — cost per
session, per model, per runtime, per skill; which tools are slow; which
sessions never made progress; what the agent was actually thinking when it
decided to rm -rf something. See the dashboard tour.
Governance. Detect an agent that has gone off the rails and — if and only if you have opted in — stop it. See Guard.
What makes it different#
It is local first. The database is a DuckDB file on your machine. There is no account required, no API key, and no network call in the default install. Cloud sync is optional, and when you turn it on, session content is encrypted with AES-256-GCM before it leaves the machine — the server stores ciphertext it cannot read. See End-to-end encryption.
It reads what already exists. There is no SDK to import and no instrumentation to add for the supported runtimes. You install it and it finds your agents. For agents that write nothing to disk — a production service built on an SDK — there are two explicit paths: the interceptor and the ingest API.
It refuses to invent numbers. Runtimes differ wildly in what they persist.
PicoClaw's message struct has no token field at all; Cursor's backend is
proprietary; Kimi records a full token breakdown but never the model id. Where
the data is not there, ClawMetry reports it as unknown and says why, rather than
showing a plausible dollar figure. A $0.00 and an unknown are different
facts and the UI keeps them different.
Severity is measured in money. A Guard incident carries
spend_at_risk_usd — the estimated cost of the flagged stretch, not the whole
session — plus the basis that produced it. Sessions are ranked by what they are
burning, because ranking a list by an invented severity score is how real
incidents end up on page three.
What it is not#
- It is not a hosted APM. Nothing is sent anywhere by default.
- It is not a prompt-evaluation product. There is an
eval suite runner, but the core is telemetry, not scoring.
- It is not a proxy you must route traffic through. The
enforcement proxy exists and is optional; observation does not need it.
- It is not, by default, able to touch your agents. Intervention is off until
you turn it on, and there are exactly four surfaces that can do it — all documented in Guard.
Who it is for#
- A developer with several agents who wants one place to see what they cost
and what they are doing.
- A team lead who needs to know which runtime, model or person is driving
spend, and wants alerts before a bill arrives rather than after.
- A platform or security engineer who has to answer "what did the agent do,
on which machine, with what permissions, and can we prove it?" — see Tamper-evident integrity and Hardening.
Next#
Install it, then walk the quickstart.