tokenclaw

Anthropic's Agent SDK billing split hits June 15. Know what your agents are spending before they stop.

On June 15, programmatic Claude usage (Agent SDK, claude -p, GitHub Actions, third-party apps) moves to a separate credit pool. When credits run out, automated requests hard-stop. No fallback, no rollover.

tokenclaw shows you both pools, side by side.

npm install -g tokenclaw
tokenclaw credits --set-pool 200    # your plan's credit allocation
tokenclaw credits

Credit Usage  (2026-06-01 -> 22d remaining)

Subscription (interactive)
  Claude Code              $4.9K consumed @ API rates
  Claude Desktop           $83.64 consumed @ API rates
  $5.0K consumed on a $200.00/mo plan = 25x value

Agent SDK credits
  ██░░░░░░░░░░░░░░░░░░░░░░░░░░░░  $13.35 / $200.00 (7%)
  ├─ OpenClaw               $13.35

Projection
  Daily rate:          $1.33/day
  Month-end forecast:  $42.71
  ✓  On track to stay within $200.00 pool

No auth, no config. Reads local session logs from Claude Code, OpenClaw, Cursor, Windsurf, Cline, Roo Code, Aider, and Continue.dev.

Credit pools (new)

tokenclaw credits                               # see both pools
tokenclaw credits --set-pool 200                # set your SDK credit allocation
tokenclaw credits --set-pool 20                 # Pro plan
tokenclaw credits --set-pool 100                # Max 5x
tokenclaw credits --set-period-start 2026-06-15 # set billing cycle start
tokenclaw credits --json                        # programmatic output

Shows subscription usage (interactive, unlimited) vs Agent SDK credits (capped, metered at API rates). Projects when you'll hit your cap based on current burn rate.

| Your plan | SDK credit pool | |-----------|----------------| | Pro ($20/mo) | $20/month | | Max 5x ($100/mo) | $100/month | | Max 20x ($200/mo) | $200/month |

30-second setup

npm install -g tokenclaw
tokenclaw today

Shows what you've spent today, by tool:

  OpenClaw           $12.41
  Claude Code        $6.20

  Total              $18.61 / $50.00 (37%)

Catch runaway costs early

tokenclaw alert --daily 50

Get notified before a bad loop burns hundreds overnight. Installs a background check (runs hourly, survives reboots on macOS).

Connect Slack to get alerts when you're not watching:

tokenclaw config slack https://hooks.slack.com/services/YOUR/WEBHOOK/URL

See cost after every Claude Code turn

tokenclaw hook install

Every time Claude finishes a turn, you see:

[tokenclaw] $18.32/$50 (37%)

Hard-block a runaway agent

tokenclaw proxy                                     # start the proxy
ANTHROPIC_BASE_URL=http://localhost:4040 claude      # point agent at it
tokenclaw cap --key sk-ant-research --daily 10      # block at $10/day

At 80% the proxy warns. At 100% it returns 429 and the agent stops. Your money stops leaving.

Built for

• Solo builders with Claude automations hitting the June 15 credit split
• Claude Code users on Max plans burning $200+/day
• OpenClaw / OpenRouter API users running cron agents
• Cursor and Windsurf power users
• Anyone who's had a surprise API bill

All commands

| Command | What it does | |---------|-------------| | tokenclaw credits | Subscription vs SDK credit usage, burn projection | | tokenclaw today | Today's spend by tool | | tokenclaw view | Full breakdown: models, projects, trends, efficiency | | tokenclaw status | One-line spend check (instant) | | tokenclaw alert --daily 50 | Alert at $50/day | | tokenclaw alert --velocity 0.50 | Alert when burning >$0.50/min | | tokenclaw hook install | Show spend after every Claude Code turn | | tokenclaw cap --key <prefix> --daily <n> | Hard-block at $n/day (needs proxy) | | tokenclaw baseline | Your spending patterns by day-of-week | | tokenclaw list sessions | Most expensive sessions | | tokenclaw digest | Daily Slack summary | | tokenclaw mcp install | Let AI agents check their own budget |

Advanced

tokenclaw alert --velocity 0.50

Alerts on spend rate over a 30-minute window from proxy data. Catches runaway loops before they hit the total threshold.

tokenclaw baseline

After 7+ days, shows per-day-of-week patterns (median, P95). After 14 days, anomaly detection replaces the default spike threshold with one tuned to your actual spend. API-billed tools only.

tokenclaw digest --install

Sends yesterday's spend summary to Slack at 9 AM daily (macOS launchd).

tokenclaw mcp install

Gives Claude Code access to three tools: get_budget_status, get_session_cost, estimate_cost. Agents can check their budget before expensive operations and switch to cheaper models when running low.

Add to your project's CLAUDE.md:

## Cost awareness
Before expensive operations, check budget via the tokenclaw MCP server.
If remaining budget is <20%, prefer Sonnet over Opus for routine tasks.

Add to ~/.claude/settings.json:

{
  "hooks": {
    "Stop": [{ "matcher": "", "hooks": [{ "type": "command", "command": "npx tokenclaw status --oneliner 2>/dev/null || true" }] }]
  },
  "mcpServers": {
    "tokenclaw": { "command": "npx", "args": ["tokenclaw", "mcp"] }
  }
}

Caps require the proxy (tokenclaw proxy). The proxy sits between your agent and the API, counts tokens, and blocks when a cap is hit.

tokenclaw cap --key sk-abc --weekly 500
tokenclaw cap --key sk-abc --monthly 2000
tokenclaw cap                               # view all caps
tokenclaw cap --clear --key sk-abc          # remove a cap

Auto-detects provider: /v1/messages goes to Anthropic, /v1/chat/completions goes to OpenAI.

Config

tokenclaw config                            # show current config
tokenclaw config slack <url>                # set Slack webhook
tokenclaw config reset                      # reset to defaults

Uninstall

npm uninstall -g tokenclaw
rm -rf ~/.tokenclaw

License

MIT