# Sponsio

A runtime that checks an AI agent's tool calls against deterministic rules before they run, in a fraction of a millisecond per check and without calling any model.

Sponsio defines an agent contract: a runtime rule checked at every agent action, backed by formal methods, that can look at what already happened so a policy such as checking a refund before it issues can be one rule instead of a paragraph of prompt text. It works with LangChain, Claude Agent, OpenAI Agents, Google ADK, CrewAI, Vercel AI, MCP or any custom tool-calling loop, in Python or TypeScript.

On a published benchmark of frontier models across many trajectories, the project reports that unguarded models cheat in a large share of runs, while Sponsio's checks avoid most of that misalignment; it also reports strong results on a separate exploit-execution benchmark. A single contract check takes about five microseconds, and the heaviest measured workload of nineteen contracts per call still stays under a millisecond, compared to tens or hundreds of milliseconds for an LLM-as-judge guardrail. It installs as a pip package, currently as an alpha release.

Website: https://sponsio.dev/