Saffron Health just open-sourced Libretto, a toolkit that makes AI-driven browser automation actually reliable. The company built it for their own needs: maintaining integrations with healthcare software like EHR systems and payer portals that don't have proper APIs. Instead of letting an LLM loose on a webpage and hoping for the best, Libretto gives coding agents a token-efficient CLI and live browser.
The toolkit works by generating Playwright TypeScript scripts that run deterministically. You can show it a workflow once, and it converts your actions into a replayable script. Or you can point it at an existing browser automation and ask it to reverse-engineer the underlying API calls. Faster and more stable than UI automation. It supports OpenAI, Anthropic, Gemini, and Vertex for its snapshot analysis, keeping heavy visual context out of your main agent's context window.
Retriever AI's founder showed up in the Hacker News thread to compare notes. Their competing approach runs entirely inside a browser extension, so auth tokens and headers propagate automatically and scripts execute at zero token cost with full determinism. Both projects share the same thesis: runtime agentic inference is expensive and unpredictable. The future is recorded workflows converted to scripts. The difference is where the recording happens.
There's an elephant in the room, though. Reverse-engineering healthcare payer portals carries real legal risk. HIPAA violations can hit $1.5 million per category annually. The Computer Fraud and Abuse Act can treat Terms of Service violations as criminal offenses. Most payer portals explicitly ban automated access. Saffron Health is clearly betting that the need for interoperability justifies the risk, but anyone using Libretto for healthcare data should talk to a lawyer first.