Free / open source
Stop the agent. Don't just alert on it.
Polls Anthropic's real Admin Usage & Cost API once a minute and deactivates the API key the instant your spend crosses a limit you set. Not another cost dashboard — everything else on the market alerts you, killcord actually stops the agent. Free, self-hosted, open source (MIT).
Install (no npm, no account)
curl -O https://www.edgethirteen.com/killcord/killcord.js export ANTHROPIC_ADMIN_KEY=sk-ant-admin01-... export KILLCORD_WORKSPACE_ID=wrkspc_... export KILLCORD_LIMIT_USD=25 export KILLCORD_KEYS=apikey_... node killcord.js watch
Put it on a 1-minute cron for continuous coverage. Give each agent its own Anthropic Workspace (Anthropic's own recommendation for attributable, revocable spend) and this becomes an accurate, real-dollar trip wire per agent.
Multi-workspace fleets, OpenAI project-key kill support, generated systemd/launchd unit. Early-access price, delivered by email within 24 hours.
Why this exists
In 2026, unattended agent loops have produced a $6,531 AWS bill in a day, a $47,000 11-day LangChain loop, and reports that most orgs can't quickly terminate a misbehaving agent once it's running. Every "LLM cost monitoring" tool on the market alerts you. killcord polls Anthropic's real Admin Usage & Cost API once a minute, and the instant your spend crosses the line you set, it calls the Admin API again to deactivate that key — the agent's next request fails, not "we'll email you."
Frequently asked
How is this different from a cost-monitoring dashboard or Slack alert?
An alert doesn't help if you're asleep, or if the runaway loop burns hundreds of dollars in the ten minutes before you see it. killcord doesn't notify — it calls Anthropic's Admin API to flip the offending key to inactive the moment it crosses your line, the same second it checks.
Is this safe to trust with an Admin API key?
It's self-hosted and dependency-free — the whole thing is under 200 lines of plain Node, readable in five minutes. It talks straight to api.anthropic.com with the key you provide via environment variable. Nothing is sent to us, stored, or proxied through any third-party server.
Do I need npm to install it?
No. Download the single script below and run it directly with Node — no package manager, no dependencies, no account required.
What does killcord NOT do?
It doesn't proxy or rate-limit calls (see LiteLLM/Requesty for that), and it can't claw back spend that already happened — Anthropic's own docs note a small amount of usage can land before a key deactivation propagates. It's a stop-loss, not a hard cap. It only deactivates the specific key IDs you list, never your whole organization.
What's in the free tier vs. Pro?
Free covers one Anthropic workspace, one spend threshold, run on your own cron — everything most solo builders need. Pro ($15 one-time, early-access price) adds multi-workspace fleets, per-key limits, OpenAI project-key kill support, and a generated systemd/launchd unit so it survives a reboot without hand-rolled cron.
MIT licensed, single dependency-free file, ~200 lines — read it before you trust it with an admin key.