Why We Built a Daily Spend Dashboard After Token-Based Billing Broke Our Budget — And the Hard Cap We Now Enforce

TL;DR: Three tools, three billing models (Copilot AI Credits, Claude Code credit pool, Anthropic API direct). We built ccost — a unified CLI that normalizes all three into daily USD burn rate and fires Slack alerts at 80% of budget. May 2026: $287 actual. June projection without tracker: $600+. With tracker: $139 controlled. Tracker source →

The Context

Two-dev team. May 2026 stack: Copilot Business ($19/seat) + Anthropic API direct for Opus ($249). June 1: Copilot switches to token-based “AI Credits” (agentic workflows = 10–50x credit burn). June 15: Claude Code credit pool launches ($100/mo for ~100 credits/day Sonnet-equivalent). Three different token economies, zero unified visibility.

What We Tested

Approach	Coverage	Verdict	Why
Copilot dashboard only	Copilot only	❌	No API or Claude Code visibility
Anthropic Console only	API only	❌	No Copilot or Claude Code pool visibility
Custom ccost CLI	All three	✅	Normalizes credits → USD; daily alert at 80% budget
token-tracker (community)	Copilot + API	⚠️	No Claude Code pool support; unmaintained

The Pivot Point

June 3, 2026: Copilot invoice projection $412 (vs $38 prior month). Same week: Claude Code credit pool announced. Realized we had no way to answer “what’s our total AI spend today?” Built ccost that weekend. First alert fired June 10: “⚠️ Copilot credits: 3,840/5,000 (77%). Review agentic usage.” Caught a runaway Composer session before it hit budget.

What We Use Now

ccost CLI (~/.local/bin/ccost, in shared dotfiles):

#!/bin/bash
# ccost — unified AI coding spend tracker
# Usage: ccost --budget=500 --alert=80

BUDGET=${1:-500}  # monthly USD cap
ALERT_PCT=${2:-80}

# Copilot: gh api /orgs/{org}/copilot/usage -> credits_used * $0.02/credit (approx)
COPILOT_CREDITS=$(gh api /orgs/basso-digital/copilot/usage --jq '.total_credits_used // 0')
COPILOT_USD=$(echo "$COPILOT_CREDITS * 0.02" | bc -l)

# Claude Code: claude-code usage --json -> credits_used * $0.05/credit (pool rate)
CLAUDE_USAGE=$(claude-code usage --json --since "1 month ago")
CLAUDE_CREDITS=$(echo "$CLAUDE_USAGE" | jq '.credits_used // 0')
CLAUDE_USD=$(echo "$CLAUDE_CREDITS * 0.05" | bc -l)

# Anthropic API: direct from console (manual entry or API if available)
API_USD=${ANTHROPIC_API_USD:-0}

TOTAL=$(echo "$COPILOT_USD + $CLAUDE_USD + $API_USD" | bc -l)
ALERT_THRESHOLD=$(echo "$BUDGET * $ALERT_PCT / 100" | bc -l)

echo "📊 AI Coding Spend (MTD)"
echo "  Copilot:  $COPILOT_CREDITS credits  ~\$${COPILOT_USD}"
echo "  Claude:   $CLAUDE_CREDITS credits   ~\$${CLAUDE_USD}"
echo "  API:      ~\$${API_USD}"
echo "  TOTAL:    ~\$${TOTAL} / \$$BUDGET"

# Daily cron also checks:
DAILY_CLAUDE=$(claude-code usage --json --since yesterday | jq '.credits_used // 0')
if [ "$DAILY_CLAUDE" -gt 85 ]; then  # 85% of 100 pool
  slack-post "#dev-alerts" "⚠️ Claude Code daily pool: $DAILY_CLAUDE/100 credits. Opus flag review needed."
fi

# Alert logic
if (( $(echo "$TOTAL > $ALERT_THRESHOLD" | bc -l) )); then
  slack-post "#dev-alerts" "🚨 AI spend at \$$TOTAL (${ALERT_PCT}% of \$$BUDGET budget). Audit needed."
fi

Cron (runs 5pm daily + 10am daily pool check):

0 17 * * * ccost --budget=500 --alert=80 >> ~/.ccost.log 2>&1
0 10 * * * claude-code usage --json --since yesterday | jq '.credits_used' | xargs -I{} bash -c 'if [ {} -gt 85 ]; then slack-post "#dev-alerts" "⚠️ Claude pool: {}/100"; fi'

Team Protocol:

• Monthly budget: $500 (covers all three)
• Alert at 80% ($400) → mandatory usage audit
• Opus usage requires --reason "auth|infra|payments" flag (logged)
• Copilot agentic workflows → migrate to Cursor/Claude Code

When You’d Choose Differently

• Single-tool shops: Native dashboards suffice (Copilot org view, Anthropic Console).
• Enterprise with FinOps team: Vendor APIs (Copilot, Anthropic) feed into centralized cost management.
• Light users (<$50/mo): Manual monthly check fine; tracker overhead not worth it.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	4	Solved a real multi-vendor blind spot
Ease of Use	3	CLI + cron setup; not plug-and-play
Value	5	Prevented 2+ budget overruns in month 1
Support	2	Internal tool; community token-tracker abandoned

---

This is part of our AI Coding Tool Evaluation series. See tracker source: ccost: Unified AI Coding Spend CLI