/meter — API Spend Tracker

When the user invokes /meter, run a SINGLE bash command that handles everything: setup check, backfill, and summary. One prompt, one approval.

Execution

Run this single bash command:

bash "$(for d in "$HOME/.claude/skills/agent-meter" "$HOME/.claude/skills/meter" ".claude/skills/meter" ".claude/skills/agent-meter"; do [ -f "$d/meter.sh" ] && echo "$d/meter.sh" && break; done)"

If meter.sh is not found, tell the user the script wasn't found and suggest reinstalling.

Format the output as a clean markdown summary for the user.

If the user passes arguments like /meter --by model or /meter --last 7d, append them: bash .../meter.sh --last 7d

Setup

Install from ClawHub into your project:

clawhub install agent-meter --dir .claude/skills

Then run /meter — it handles everything automatically: - Copies the session-end hook to .claude/hooks/ - Creates or updates .claude/settings.json with the Stop hook - Backfills any existing Claude Code sessions - Shows your spend summary

That's it. Future sessions are tracked automatically via the Stop hook.

Dashboard sync (optional)

To sync spend data to the hosted dashboard at dashboard.agentmeter.io:

bash .claude/skills/agent-meter/meter-sync.sh --setup

The setup wizard will prompt for your API key and agent name. Get a key from the dashboard's Machines page. After setup, sync anytime with:

bash .claude/skills/agent-meter/meter-sync.sh

What the scripts do

meter.sh — Single entry point for /meter. Checks hook installation, backfills if needed, shows spend summary. All in one script, one permission prompt.
meter-session-end.sh (Stop hook): Parses the current session's transcript for per-message token usage. Detects model (opus/sonnet/haiku), calculates cost with model-specific pricing including cache tier pricing, writes a single session_summary record. Deduplicates by session_id — safe across session pause/resume. No network calls, no secrets access, jq dependency only.
meter-parse-claude-sessions.sh (backfill): Scans all Claude Code transcripts in ~/.claude/projects/*/*.jsonl. Extracts the same usage data as the hook. Deduplicates against existing records by session_id. Safe to re-run — only processes new sessions. No network calls, no secrets access, jq dependency only.

Query Examples

SPEND="$HOME/.agent-meter/spend.jsonl"

# Spend by project
jq -s 'group_by(.project) | map({project: .[0].project, cost: (map(.cost_usd) | add), sessions: length}) | sort_by(-.cost)' "$SPEND"

# Spend by model
jq -s 'group_by(.model) | map({model: .[0].model, cost: (map(.cost_usd) | add), sessions: length}) | sort_by(-.cost)' "$SPEND"

# Today's spend
jq -s "[.[] | select(.ts | startswith("$(date -u +%Y-%m-%d)"))] | map(.cost_usd) | add" "$SPEND"

# Top 10 most expensive sessions
jq -s 'sort_by(-.cost_usd) | .[:10] | .[] | "(.cost_usd | . * 100 | round / 100) (.project) (.total_calls) calls (.ts | split("T")[0])"' "$SPEND" -r

Schema

All records are session_summary type in ~/.agent-meter/spend.jsonl.

Field	Required	Description
`type`	yes	Always `session_summary`
`ts`	yes	ISO 8601 timestamp (session start)
`api`	yes	`api.anthropic.com`
`model`	yes	Model identifier (e.g. `claude-opus-4-6`)
`session_id`	yes	Claude Code session UUID
`project`	yes	Project directory name
`total_calls`	yes	Number of assistant messages
`tokens_in`	yes	Input tokens (non-cached)
`tokens_out`	yes	Output tokens
`cache_creation`	yes	Cache creation input tokens
`cache_read`	yes	Cache read input tokens
`cost_usd`	yes	Estimated cost (model-aware pricing)
`source`	yes	`hook` or `session_parse`
`purpose`	no	Human-readable intent (manual tag)
`intent`	no	Machine tags (manual)

Cost Calculation

Per-1M-token pricing:

Model	Input	Output	Cache Create	Cache Read
Opus	$15	$75	$18.75 (1.25x)	$1.50 (0.1x)
Sonnet	$3	$15	$3.75 (1.25x)	$0.30 (0.1x)
Haiku	$0.25	$1.25	$0.3125 (1.25x)	$0.025 (0.1x)