VAIL Agent Assurance
VAIL Agent Assurance helps you check the AI agents you build. It reads an agent's own definition (its instructions, skill files, tool schemas and code) and works out how the agent could depart from what it was told to do. What you can do: - Upload an agent's source and scan it. The scan returns behavior risks, the ways the agent could do something unintended, and control behaviors, the things it must keep doing correctly. Each one cites the files and tools in the source behind it. - Generate probes: test messages with deterministic checks that score the agent's reply, written for each behavior. - Retrieve behaviors and probes for any agent in your organization, and follow scans in progress. - Download VAIL's grading package to run probes against an agent on your own machine, graded the way VAIL grades them. Run the assurance loop with your coding agent to harden the agent before you deploy. Test agents with different models to see how behavior of the agent changes, improves, or degrades with varying levels of intelligence. Works with agents built on LangGraph, Hermes, OpenClaw and other frameworks, or any agent reachable over HTTP. You need a VAIL account, and you sign in through your browser. Scanning sends the agent source you choose to VAIL, which stores it for that agent. Uploading source from your machine needs a coding agent with shell access, such as Claude Code. In chat, you can review results and probes for agents already scanned.
Details
https://agents.projectvail.com/mcp · streamable-http
