Libretto gives a coding agent a live browser and a command-line interface built for low context use, so the agent can build and maintain web integrations. It inspects live pages, captures network traffic to work out a site's underlying API, records user actions and replays them as automation scripts, and debugs a broken workflow interactively against the real site.
It is used mainly as a skill invoked from a coding agent: a prompt can ask it to scrape a site one-shot, to watch you perform a workflow in the browser and turn it into a script that takes named inputs, to convert an existing browser script into direct network calls, or to reproduce and fix a broken selector in an existing integration. Converting a workflow to network requests can be paired with an analysis of the captured requests for common security cookies, to judge whether the API path is safe to use.
The same operations are available directly from the command line: open a URL, run a workflow headless, capture a screenshot with a compact accessibility tree, execute TypeScript against the open page, and close the browser. Sessions are named, and failed or paused workflows keep the browser open so the exact page state can be inspected before a fix or resume; successful runs close by default unless told to stay open. All state lives in a .libretto/ directory at the project root, and setup installs agent skills and downloads Chromium.
The command-line interface records telemetry limited to an install id, timestamp, command event name, an error flag, package version, build channel, and a cloud user id when signed in — not arguments, URLs, project paths, cookies, keys, error details, or emails. It can be disabled with an environment variable, a do-not-track setting, a CI flag, or the local telemetry file.
Features
- Live page inspection: read pages with minimal context overhead through snapshots and a compact accessibility tree
- Network capture: record requests to reverse-engineer a site's API and generate direct network scripts
- Action recording: perform a workflow by hand and have it rebuilt as a script with named inputs
- Interactive debugging: reproduce a failure, pause at any point, inspect the live page, and fix the workflow
- Session-scoped commands: every command accepts --session to target a specific browser session
- Headless runs: run a workflow closed-on-success, or keep a successful run inspectable
- Project-local state: configuration, sessions, and profiles live under .libretto/
- Telemetry controls: opt out through LIBRETTOTELEMETRYDISABLED, DONOTTRACK, CI=1, or the local telemetry file
