A new browser built for machines. Not a fork. Not a patch.
Come chat: discord.gg/K63XeymfB5
Browsers waste time loading scripts one at a time.
We shipped <link rel=preload as=script> support so Lightpanda now starts downloading blocking scripts up front instead of as the parser hits them.
Same page, same network, six scripts, each delayed 300ms server-side:
Old build: 1813ms
New build: 606ms
https://gist.github.com/katie-lpd/aea8c50b37f104a010d0c48897b5e9d7
File uploads now work in Lightpanda! 🗃️
Attach a file to an input[type=file], submit, and Lightpanda encodes it as multipart/form-data, same as a real browser.
github.com/lightpanda-io/demo/blob/main/puppeteer/form_file.js
https://github.com/lightpanda-io/demo/blob/main/public/form/input_file.html
Headless browsing for agents lives or dies on knowing when a page is actually ready.
Lightpanda now supports networkalmostidle as a --wait-until condition, so your agent moves the moment the page is effectively loaded.
Fix 3: an integrated agent that caches the system prompt.
MCP pays round-trip overhead and re-reads big prefixes at full price every call. Cached input is ~10x cheaper on Anthropic's pricing.
Our early runs hit 99% cache reads after turn one
Fix 2: a real search tool. Tavily with
Duckduckgo fallback
The model used to fake search by loading a search engine and parsing the html by hand. A dedicated search call drops eval usage from 17% to 3% in testing
We benchmarked Lightpanda for agents and learned the tool surface you hand the model matters as much as the engine underneath it.
Here's what we're shipping because of it 🧵
https://lightpanda.io/blog/posts/benchmarking-lightpanda-for-agents
