claude-spend

The user invoked /claude-spend (or asked a natural-language question about Anthropic spend). Argument: $ARGUMENTS (empty → default 30 days).

This is the authoritative billed cost — what Anthropic actually charges — distinct from /claude-usage, which is the client-side ccusage estimate (good for breakdowns by project/session, but not the source of truth for spend).

Run

!`python "${CLAUDE_PLUGIN_ROOT}/scripts/anthropic-spend.py" $ARGUMENTS`

The script reads its Admin API key from ~/.claude/.anthropic-admin-key (single line, no quotes). If the key file is missing or empty, the script exits with setup instructions — relay them and stop.

If python is not on PATH (common on Windows when only the py launcher is installed), retry with py in place of python.

Output handling

The script's output is already a compact report: total, per-day with bar chart, per-model breakdown, link to the Console for the full GUI. Forward it verbatim — no rewriting.

Then add one short framing line at the top (or repeat it as a footer) the first time it runs in a conversation:

Source: Anthropic Admin API cost_report. This is billed cost, not the ccusage estimate.

Skip the framing line on subsequent runs in the same conversation.

Errors

• 401 / 403 — key invalid or not an Admin Key. The script's stderr already names the fix; relay it.
• 429 — rate-limited. Suggest waiting a minute, don't retry automatically.
• Network failure — relay the error, suggest checking VPN / corporate proxy.
• days out of range — the script errors clearly; relay.

Notes

• Daily buckets are UTC; the day boundary is midnight UTC, not local time.
• Cache reads count as zero-cost on Anthropic's end, so a heavy cache-read session shows lower spend here than tokens would suggest — that's expected, it's what you were billed.
• For per-project or per-session breakdowns, use /claude-usage project / session — those are estimates but they can attribute spend to specific work.