Text: We previously posted Vectorly (https://news.ycombinator.com/item?id=46209538), our open-source app that reverse-engineers websites into reusable APIs. The core idea: instead of clicking through UIs like a human, interact with websites through their underlying HTTP endpoints. Faster, cheaper, more reliable.

Building routines by recording browser sessions worked well, but refining them was still manual. So we built Guide, a conversational agent that walks you through the whole process:

- Describe your task and the agent helps scope it

- Record in a cloud browser, just use the site normally

- Agent analyzes the session, identifies backend endpoints, extracts parameters, builds a reusable routine

- Refine through chat to propose edits, review diffs, iterate without hand-editing anything

The prevailing approach to web automation is computer use agents that click and type like humans. We think this is backwards. Web app implementations are wildly diverse, but they all talk to structured backend APIs. LLMs are great at reading code and inferring structure, so we use them at build time to reverse-engineer those APIs, then the resulting routines run without AI in the loop.

We'd love to hear what you think, and happy to answer any questions about the approach! :)

- GitHub: https://github.com/VectorlyApp/bluebox-sdk

- Blog post with more detail: https://vectorly.app/blog/introducing-guide-agent

- Try it: https://console.vectorly.app/guide