--- title: Quickstart description: From pip install to reading your first agent session, answering your first cost question, and letting your coding agent query its own history. keywords: ClawMetry quickstart, agent observability tutorial, track Claude Code usage eyebrow: Get started --- # Quickstart Ten minutes, five steps, no configuration. By the end you will have the dashboard running, know what your agents cost, and have Claude Code able to query its own history. ## 1. Install and onboard ```bash pip install clawmetry clawmetry onboard --local ``` `--local` skips every account prompt and writes a marker so nothing ever leaves the machine. Drop it if you want the cloud dashboard and phone approvals; you can [connect later](/docs/cli/connect/) either way. The wizard starts the sync daemon and opens . ## 2. Confirm your agents were found ```bash clawmetry runtimes ``` You should see the runtimes that are actually installed on this machine marked as detected. Detection is a filesystem check — it looks for each runtime's real store, not for a binary on your `PATH`. ```bash title="Machine-readable" clawmetry runtimes --json | jq '.runtimes[] | select(.detected) | .id' ``` If a runtime you use is missing, its data directory has probably been relocated. Every runtime's [reference page](/docs/runtimes/overview/) lists the environment variable that points ClawMetry at it — for example `CODEX_HOME` for Codex or `CLAWMETRY_CURSOR_DB` for Cursor. :::note Give it a moment The daemon ingests on a cycle, and the first pass over a long history takes longer than the ones after it. `clawmetry status` shows how far it has got. ::: ## 3. Read a session Open the dashboard and go to **Conversations**. Pick a session and you get the transcript with thinking separated from replies, every tool call with its arguments and result, and per-turn tokens and cost. The same thing from the command line, because the store is queryable: ```bash # five most recent sessions curl -s 'http://localhost:8900/api/local/sessions?limit=5' | jq '.rows' # everything that happened in one of them curl -s 'http://localhost:8900/api/local/transcript/' | jq '.rows[] | {ts, event_type, tool_name}' ``` ## 4. Answer a cost question Go to **Cost**. The first question most people have is "what did this week cost, and where did it go" — that is the runtime and model breakdown at the top of the page. ```bash title="Same answer, in the shell" curl -s 'http://localhost:8900/api/local/aggregates' | jq '.rows[-7:]' ``` Cost is derived at ingest from the token split and the model, using a multi-provider pricing table. Where a runtime records real dollars — Cline, opencode, Pi, Copilot, Exo — ClawMetry uses those instead of estimating. Where neither is available it says so. [How cost is computed](/docs/cost/model/) explains which is which. ## 5. Let your agent query its own history This is the part people underestimate. Start the MCP server: ```bash clawmetry mcp ``` Then register it with Claude Code: ```bash claude mcp add clawmetry -- clawmetry mcp ``` Now you can ask, inside Claude Code: > Using clawmetry, what did I spend on Cursor versus Claude Code this week, and > which of my sessions had the most repeated tool failures? The agent calls `get_cost_summary` and `list_sessions`, reads its own telemetry, and answers. [Agents & MCP](/docs/mcp/overview/) covers the full tool surface and a set of recipes worth copying. ## Where to go next

What the daemon does, and why handlers never open DuckDB directly.

Every page, what question it answers, and where its data comes from.

Turn on Guard

Catch a looping agent, and decide whether ClawMetry may stop it.

Your first week

A short sequence: budgets, alerts, retention, then policies.