Zatanna Watches You Use Legacy Software Once and Turns It Into an API

The Macro: Legacy Software Does Not Have APIs and Never Will

Every large organization runs critical processes through software that has no API. Insurance claims processed through 20-year-old desktop applications. Government forms submitted through web portals built in 2005. Medical records entered into systems that predate the iPhone. These applications work, they are deeply embedded in business workflows, and nobody is going to rewrite them.

The AI agent boom has made this problem acute. AI agents need APIs to act on behalf of users. They cannot click through a legacy web form or navigate a Windows desktop application. The standard workaround is RPA (robotic process automation) tools like UiPath and Automation Anywhere that replay mouse clicks and keystrokes. But RPA is brittle. Change a button position or add a loading screen and the bot breaks.

The fundamental issue with click-based automation is that it operates at the UI layer, which is the least stable layer of any application. The underlying system calls, HTTP requests, and database operations that the UI triggers are much more stable. If you could capture those underlying operations instead of the clicks, you would have a much more reliable automation.

Zatanna, backed by Y Combinator, does exactly this. They observe a human performing a workflow, reconstruct the underlying request behavior rather than replaying clicks, and expose it as a clean API endpoint.

The Micro: Requests, Not Clicks

Rithvik Vanga (CEO), Alex Blackwell (CTO), and Tarun Vedula (COO) built Zatanna around the insight that software workflows have a request layer beneath the UI layer. When you fill out a form and click submit, the browser sends an HTTP request. That request is the real action. The form is just the interface.

Zatanna captures these request flows during a single human demonstration, then reconstructs them into a repeatable API call. It handles authentication, retries, and anti-bot detection underneath. The result is a production-grade endpoint that AI agents or internal systems can call millions of times.

This is technically harder than it sounds. Modern web applications use complex authentication flows, CSRF tokens, session management, and rate limiting. Reconstructing a reliable API from observed request behavior requires understanding all of these mechanisms and handling them programmatically.

The company already supports millions of requests for customers including Pikkit, Fleetline, CrowdVolt, and PartBay. This early traction suggests the approach works in production, not just in demos.

Competitors include UiPath and Automation Anywhere for RPA, Browserbase for browser automation, and various web scraping platforms. The key differentiator is reliability: by working at the request layer instead of the UI layer, Zatanna’s automations should be significantly more resistant to UI changes.

The Verdict

Zatanna is solving the right problem with the right technical approach. Legacy software integration is one of the biggest blockers for AI agent deployment in enterprises.

At 30 days: what percentage of workflows captured from a single demonstration work reliably in production without manual tuning?

At 60 days: how does Zatanna handle applications that change their request structure (new API versions, authentication changes)?

At 90 days: are enterprise customers using Zatanna to connect AI agents to legacy systems at scale?

I think Zatanna has a strong technical insight. Request-level automation is fundamentally more reliable than UI-level automation. If they can handle the edge cases, authentication complexity, and anti-bot measures that make this hard in practice, they become the bridge between the AI agent future and the legacy software present.