Show HN: Libretto - Making AI browser automations deterministic
Key point
Libretto was released to improve the reproducibility and debugging of browser automation.
Details
Libretto is a toolkit that helps coding agents build more deterministic and token-efficient web integrations while still working with real browsers.
Its core features are as follows.
- live page inspection: check the real page with minimal context
- network traffic capture: collect requests so you can reverse-engineer a site's API
- user action recording/replay: record actions a human performed and replay them as an automation script
- interactive debugging: directly trace and fix broken workflows on the real site
Saffron Health built it to maintain its medical software integrations, and has now released it publicly for other teams facing the same problem.
The installation flow is also straightforward.
npm install librettonpx libretto setupdownloads Chromium, installs the skill, and sets up the default snapshot modelnpx libretto statuschecks workspace readinessnpx libretto ai configure <openai | anthropic | gemini | vertex>to manually change the snapshot analysis model
There are roughly four use cases.
- one-shot script generation: for example, having the agent directly generate a script that collects the first 10 posts on LinkedIn
- interactive script building: a human performs the workflow directly in the browser, and this is used to reconstruct a Playwright script
- browser automation → network requests: replacing UI automation with direct API calls for faster, more reliable execution
- broken integration fix: reproducing issues like selector errors in the browser and fixing them while viewing the live page
A CLI is also provided.
- Commands like
open,snapshot,exec,run,resume,pages,save,closecontrol sessions - All commands can target a specific session with
--session - State is stored under
.libretto/at the project root, and configuration is kept in.libretto/config.json
The key point is that this is a tool that turns browser-based AI automation from simple 'clicking around' into a reproducible and debuggable workflow.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.