Anthropic's billing split hits June 15. Know what your agents spend.
On June 15, programmatic Claude usage moves to a separate credit pool. When credits run out, your automations stop. tokenclaw shows both pools side by side, projects when you'll hit the cap, and alerts before it happens.
$ npm install -g tokenclaw
"It's a massive bait & switch. I can't imagine how many hours were spent building on the premise of subscription pricing." -- HN, June 2026
"Agent entered an infinite loop. $4,200 in a weekend."
"I left an agent running overnight. $280. No warning."
June 15: What changes
Anthropic is splitting Agent SDK usage (claude -p, GitHub Actions, third-party apps) into a separate monthly credit pool. Pro gets $20, Max 5x gets $100, Max 20x gets $200. Billed at API rates. No rollover. When credits run out, automated requests hard-stop.
Interactive Claude (chat, Claude Code terminal) stays unlimited on your subscription. But if you run cron jobs, CI pipelines, or agent workflows through the SDK, you need to know which pool they draw from.
$ tokenclaw credits
Credit Usage (2026-06-01 -> 22d remaining)
Subscription (interactive)
Claude Code $4.9K consumed @ API rates $5.0K on a $200/mo plan = 25x value
Projection
Daily rate: $1.33/day
Month-end forecast: $42.71 ✓ On track to stay within $200.00 pool
How it works
tokenclaw sits between your agents and model providers. Every request: identify the API key, check spend against your rules, forward or block. Auto-detects Anthropic and OpenAI from the request path. One proxy.
Per-key budgets
Daily, weekly, or monthly. Each key gets its own rules.
Progressive rules
Warn at 50%, warn again at 80%, block at 100%. You decide.
Slack alerts
Get notified when thresholds fire. Or just watch the logs.
Safe by default
Nothing blocked unless you add a --block rule. Default is warn-only.
Multi-provider
Anthropic, OpenAI, Groq, Together, Fireworks. One proxy.
Runs on localhost
No cloud. No telemetry. No account. Your requests stay yours.
