Lightpanda saves an agent’s browser work as a script that needs no model to run
Francis Bouvier’s team separates discovering a workflow from repeating it, with native browser calls and explicit handling for credentials.

Lightpanda’s example script grabs three story links, visits their discussion pages and collects comments. A loop follows the links; a catch block handles missing comments. The result is something a developer can inspect after the agent session, right down to which elements supplied each field.
“You pay the model once to write the file,” cofounder Francis Bouvier writes in Lightpanda’s introduction. His team puts the browser, runtime and agent in one binary. The agent calls the engine directly, avoiding the separate browser process and Chrome DevTools Protocol messaging used by many automation clients. Lightpanda still offers CDP through its server for existing tools.
During an agent session, /save exports the work. The launch post calls this an alpha feature: with a model connected, the agent generates a script from the session’s intent; without one, it transcribes recorded commands. That distinction makes the exported file worth reading. Generating a reusable procedure is a different promise from recording each exploratory click, including the wrong ones.
The current PandaScript documentation uses lightpanda run script.js for replay. The script gets normal JavaScript variables and loops, plus a Page object for navigation, clicking and extraction. It runs in a separate V8 context from the page. A page’s document is available through explicit evaluation inside that page; it does not become the script’s global environment.
That environment also leaves out Node’s filesystem and package-loading APIs. A saved workflow cannot assume an npm dependency is waiting for it. Navigation needs await; the other page methods run synchronously. Tool failures throw exceptions, so a script can catch a missing optional section while allowing a genuinely broken step to stop the run. These details determine whether the file is useful automation or just a transcript with a JavaScript extension.
Credentials need their own treatment. The launch post says secrets become $LP_* placeholders resolved at runtime. The agent documentation also distinguishes direct slash commands from commands that require a model. Running in --no-llm mode does not suddenly give ordinary English instructions a parser. The model-free part starts with a procedure the runtime already understands, much as a player piano still needs the right roll.
The maker’s README reports 123 MB peak memory against Chrome’s 2 GB and five seconds against 46. Its benchmark methodology gives a revealing detail: the crawler visits 933 URLs, and concurrency changes both browsers’ results. At 25 Lightpanda processes it records 4.81 seconds and 123 MB; Chrome’s 25-tab row records 46.70 seconds and 2 GB. A login workflow would need its own measurement, including the time spent authenticating and checking the data it returns.
A useful next comparison would include the returned data and a deliberately changed page. If replay fails clearly when an expected field disappears, someone can repair the script before another scheduled run. Whether that failure is obvious enough to trust unattended remains the more interesting test.