FiGuard

FiGuard provides runtime authorization for AI agents, enabling pre-flight spend checks and resource control before any action is executed. The tool addresses the issue of agents overspending or exceeding resource limits through retries, loops, or parallel execution, which can go unnoticed until after the fact. By requiring agents to request authorization before performing actions that consume bounded resources—such as money, tokens, API calls, or GPU hours—FiGuard ensures that each decision is checked against defined limits, categories, and allocation rules.

The platform maintains a full audit trail of every authorization decision, including agent identity, delegation hierarchy, and the reason for approval or denial. Its spend tree structure allows for granular tracking of resource usage at every level of delegation. FiGuard logs all actions in a ledger, providing real-time updates on balances and clear reason codes for denied requests. The tool supports concurrent authorization, handles idempotency to prevent duplicate reservations, and separates reservation from confirmation to avoid race conditions and over-allocation. Security is reinforced through session tokens that are scoped to individual budgets and never stored in plaintext.

FiGuard is framework-agnostic and works alongside agent orchestration tools such as LangChain, CrewAI, OpenAI Agents SDK, Anthropic SDK, and LangGraph. It can be integrated into Python scripts, background jobs, serverless functions, or any runtime where agents interact with APIs or services that incur costs or consume limited resources. The tool is self-hostable and can run locally by default, using an in-process SQLite database without requiring a server or signup. For shared multi-agent budgets, it can be run as a server. There is also a hosted sandbox demo available for trial purposes, though it comes without service guarantees and periodic data wipes. FiGuard offers both Python and TypeScript SDKs for integration.

As a financial-grade authorization infrastructure, FiGuard brings payment industry primitives—such as pessimistic write locks, reserve-then-capture models, and secure session token handling—to agent-based systems, ensuring that resource consumption is always authorized, logged, and explainable before execution.