Skip to content
AnglinAI Docs

Anthropic Spend Kill-Switch

Unlike GCP (where gcloud billing projects unlink is a one-call hard cap), Anthropic does not expose a programmatic “detach billing” API. The kill-switch is therefore a combination of:

  1. Hard monthly spend limit configured in the Anthropic Console — Anthropic stops accepting API calls when the limit is reached. This is the actual hard cap.
  2. Usage alerts at 50% / 80% / 100% — emailed to larry@anglin.com.
  3. Application-layer safeguards (rate limit, Turnstile, history caps) that bound the worst-case burn rate while the hard cap is in flight.

The application-layer safeguards live in workers/api/src/routes/support-chat.ts.

One-time setup in the Anthropic Console

  1. https://console.anthropic.com/settings/billing → Cost section
  2. Set Monthly spend limit = $200 (initial value — bump as legitimate usage warrants)
  3. Set Email alerts at 50%, 80%, 100% to larry@anglin.com
  4. Confirm the affected key (ANTHROPIC_API_KEY, used by the support chat) is on this workspace. Each Anthropic workspace has its own limit.

Once the monthly limit is hit, every messages.create returns HTTP 429 / billing_limit_exceeded until the next calendar month or until the limit is raised.

What happens when the cap fires

  • Support chat at https://theaccessible.org/support starts returning the generic “Assistant error” message — clients fall through to the contact form (intended behavior).
  • All other Anthropic-using surfaces fail simultaneously. Today the support chat is the only public consumer; if you add more, list them here so on-call knows the blast radius.

Investigating before raising the cap

  1. Read recent abuse logs in Supabase: 
    SELECT level, message, metadata, created_at
    FROM app_logs
    WHERE service = 'support-chat'
      AND created_at > now() - interval '24 hours'
    ORDER BY created_at DESC
    LIMIT 200;
  2. Group by metadata->>'ip' and metadata->>'status' to spot rate-limited or turnstile-failed bursts: 
    SELECT
      metadata->>'ip'      AS ip,
      metadata->>'status'  AS status,
      count(*)             AS n
    FROM app_logs
    WHERE service = 'support-chat' AND created_at > now() - interval '24 hours'
    GROUP BY 1, 2 ORDER BY n DESC LIMIT 50;
  3. Check the Anthropic Console usage chart at https://console.anthropic.com/settings/usage — confirm the shape (steady growth = legit, vertical wall = abuse).

If abuse is confirmed:

  • Rotate ANTHROPIC_API_KEY (see docs/admin/secrets-inventory-and-rotation.md).
  • Add the offending IP/CIDR to a Cloudflare WAF rule before re-raising the cap.
  • workers/api/src/routes/support-chat.ts — rate limits, Turnstile, history caps, abuse logging
  • docs/admin/gcp-budget-killswitch.md — sibling document for Gemini spend
  • GitHub Issue #711 — original tracking issue for support-chat safeguards