Pydantic AI Shields — Guardrails for Pydantic AI

Main

• Budgeting by default: add CostTracking(budget_usd=...) to stop runaway agent loops and to record total tokens/cost per run.

• Tool permissions: use ToolGuard(blocked=[...], require_approval=[...]) so unsafe tools never appear (or require explicit approval).

• Input/output controls: InputGuard blocks risky user prompts early; OutputGuard can enforce redaction or policy checks post-run.

• Prefer incremental rollout: start with cost tracking + tool allowlist, then add input/output guards for the highest-risk surfaces.

Source-backed notes

• README shows pip install pydantic-ai-shields and a Quick Start example using CostTracking, ToolGuard, and InputGuard capabilities.
• README describes CostTracking as tracking tokens/cost with optional budget enforcement and raising BudgetExceededError.
• README explains ToolGuard supports blocking tools entirely and requiring approvals via a callback.

FAQ

• Is this a full agent framework?: No — README positions it as drop-in capabilities for Pydantic AI, not a separate agent runtime.
• Can I block a tool completely?: Yes — README shows ToolGuard(blocked=[...]) removes tools from the model’s tool list.
• What’s the first guardrail to add?: Cost tracking + tool permissions; then add input/output guards for your highest-risk prompts.