A single recursive or retry loop can call your model thousands of times in an hour, on your key, while you sleep. A dashboard tells you it happened. AgentGuard stops the call before it runs: a hard spend cap and a kill switch, enforced inside your own process, before the provider is ever charged.
Helicone, Langfuse, and provider dashboards are good at one thing: telling you what already happened. By the time the alert fires, the tokens are spent and the bill is real. A spend cap is a different job. It has to refuse the call at the moment it would exceed your limit, in the same process, before any money moves. That is a control, not a chart.
Per-call, per-day, and per-month ceilings. At the limit the call is blocked before the provider is touched, so a runaway loop stops itself.
Flip it and every model call stops immediately. The panic button you wish you had at 3am.
A read-only agent physically cannot call a money-moving tool. Scope what each agent is even allowed to attempt.
Each decision signs an Ed25519, content-free receipt anyone can verify. "We cap our agents" becomes something you can prove.
It wraps your existing OpenAI, Anthropic, or OpenRouter client. No proxy, nothing to security-review, prompts never leave your process.
import { withSpendGuard } from '@agentguard-run/spend'; const guarded = withSpendGuard(new Anthropic(), { policy: { caps: [{ amountCents: 2000, window: 'per_day', action: 'block' }] }, scope: { tenantId: 'acme', agentId: 'my-agent' }, }); // over $20/day? the call throws before it ever reaches the provider
On LangChain, LangGraph, CrewAI, or the Vercel AI SDK? See the framework drop-ins. Want us to handle AI access with no key to manage? That is Managed, $99/mo.
The free scan shows your governance score, the specific gaps, and what a runaway loop could cost you. No signup, nothing to install.
Scan your agent, free →