What is AgentGuard Spend?

AgentGuard Spend is an MIT-licensed SDK that wraps your LLM provider client (OpenAI, Anthropic, AWS Bedrock) and blocks runaway AI agent spend before the provider is ever called. Every decision is Ed25519-signed and hash-chained. Available on npm as @agentguard-run/spend and PyPI as agentguard-spend.

How does AgentGuard differ from LiteLLM or Portkey?

Proxy gateways like LiteLLM and Portkey sit in the network path — your prompts and API keys flow through their servers. AgentGuard runs in-process inside your runtime, so prompts and keys never leave your environment. Zero data plane involvement. For regulated or sovereign deployments, the proxy approach is disqualified.

Can AgentGuard block calls before the provider is charged?

Yes. Cost is projected from the prompt and max_tokens before dispatch. If any active cap (per-call, per-minute, per-hour, per-day, per-month) would be exceeded, an AgentGuardBlockedError is thrown before the provider receives the request. The provider never sees the call. You never get the bill.

What providers does AgentGuard work with?

Any provider where you bring your own client: OpenAI, Anthropic, AWS Bedrock, plus framework integrations for LangChain, LangGraph, CrewAI, LlamaIndex, Hermes, OpenClaw, and the Vercel AI SDK.

Is AgentGuard patent-protected?

Yes. AgentGuard Spend is protected by five US provisional patent applications (63/983,615; 63/983,621; 63/983,843; 63/984,626; 64/071,781) filed February-May 2026, all assigned to Dunecrest Ventures Inc. with assignments recorded at the USPTO.

Spend governance
that never sleeps.

AgentGuard Spend wraps your provider SDK and blocks runaway AI agent spend before the provider charges you. 30-second install. Your keys never leave your runtime. Every decision is signed and verifiable.

Node.js

npm install @agentguard-run/spend

Python

pip install agentguard-spend

Start here

Quickstart →

Install, wrap your provider client, set a daily cap. 60 seconds.

Capabilities →

Every cap window, action, scope dimension, capability tier, and locale.

Workflows →

One budget envelope, resumable checkpoints, and a signed chain for multi-step runs.

Integrations →