Open Source Radar — October 31: Lightpanda, a Browser for Agents
Today's board is plumbing: a browser with no rendering engine, a governance gateway for coworker agents, and two official skill libraries — the layer agents actually run on.
Lightpanda (Zig, 38,722 stars, AGPL-3.0) is a headless browser written from scratch for agents instead of another Chromium fork, and it took a daily trending slot again today with a push this morning. Its own benchmark — 933 real pages on an AWS EC2 m5.large — puts peak memory at 123 MB for 100 pages against roughly 2 GB for headless Chrome, and 5 seconds against 46, about 16x less memory and 9x faster. The new part is that it now ships an agent of its own: describe a task in plain English, it drives the browser in the same process, then exports the session as a deterministic JavaScript script you can replay later with no model in the loop. There is a native MCP server as well, with one browsing session per connection so two agents stop clobbering each other's page. The trade-off is architectural — it has no rendering engine, so you get text, markdown and structured data rather than pixels, and its PNG and PDF output are text-only renders. Use it for scraping, form-filling and crawling at scale; do not buy it for visual QA.
CopilotKit OpenBot (TypeScript, 9,009 stars, MIT) is the second repo named OpenBot to trend this week, and the same idea from the opposite end: a chat workspace of "coworkers", each with a container of its own — Chromium, files, saved logins — running entirely on your infrastructure. What separates it from the nightly-labs project is the gate. Every browser, file, shell and MCP call resolves through one gateway that evaluates a policy, writes an audit row, and only then touches the computer; deny is checked before allow, a missing policy permits nothing, and a malformed rule refuses rather than opens. Any AG-UI endpoint becomes a coworker, so LangGraph, CrewAI, Pydantic AI and hand-written agents all arrive the same way, and a person can take the wheel at a login wall and hand it back. Two caveats from the README: it is beta, and it is not standalone — durable threads and memory need CopilotKit Intelligence, whose free Developer plan will only run a local 30-day preview on a Mac with Docker.
anthropics/skills (Python, 185,619 stars, 21,914 forks) is the reference shelf every other skills repo in this post is copying: SKILL.md folders with YAML frontmatter, the Agent Skills specification itself, a template, and example skills across creative work, testing, MCP server generation and enterprise workflows. Two things are worth knowing before you fork it. The document skills that power Claude's own file handling — Word, PDF, slides, spreadsheets — are technically source-available rather than open source, included deliberately as a look at production-grade skills; most of the rest is Apache 2.0. And 1,451 open issues against 185,000 stars is the honest ratio: this is a demonstration repository by its own description, so behaviours you get from Claude may differ from what the files show.
expo/skills (Shell, MIT, 3,961 stars) is the clearest example yet of a framework vendor shipping documentation as agent skills: roughly two dozen of them, covering project structure, Expo Router, animation, native UI, design tokens, native modules, brownfield adoption and SDK upgrades, installable into Claude Code, Codex, Cursor, Copilot, Windsurf, Gemini, Cline and Kiro. The design decision worth stealing is honesty about money — the skills are split into a free open-source framework group and a services group whose whole purpose is paid Expo Application Services, and every services skill opens with a note on costs and plan limits. It also bundles Expo's MCP server, so the agent can read live docs, kick off builds and screenshot a running simulator.
Worth watching this week: whether the Agent Skills spec gets versioned releases with the repo, and whether other framework vendors copy Expo's free-versus-paid split instead of shipping one undifferentiated pile.
Which of these would you actually let near a logged-in browser? Write to us via our contact page.
