Tula is an open-source personal health AI built as a set of OpenClaw skills. In this live technical session, we will inspect the actual skill structure, run validation checks, execute behavioral evals, and show how a self-hosted agent can be owned, tested, and improved without disappearing into a vendor black box.

The demo moves through three layers, the OpenClaw runtime, the Tula health skills, and the Waza eval harness. We will look at real SKILL.md files, run Waza checks against the repo, execute behavioral tests for health-record and medical-document workflows, and write one new test from a plausible failure mode: “the agent tries to move protected health information outside the workspace.”

The key idea is simple. Owning your agent does not just mean hosting it yourself. It means owning the runtime, the skills, the data boundary, and the tests that prove the behavior still holds after the next change.

Attendees will leave with a practical pattern for building agent systems that are open, inspectable, and verifiable.