Open Source Radar — November 7: LobeHub's 84,000-Star Super-Agent
Today's board is about agents that hold down a job rather than answer a prompt: a team-chat super-agent with 84,000 stars, a Rust engine that studies any app so an agent can rebuild it, and a skill that strips the AI look out of your interface.
LobeHub (TypeScript, about 84,300 stars) is the open-source platform behind LobeChat, rebuilt around what it calls super-agents: you add one to the team chat you already use, plug in any model or IM app, and it runs several tasks at once, picks up work nobody handed it, and accumulates expertise the longer it runs. The reason it is moving this week is a hands-on test from Chinese outlet XinZhiYuan, which built a three-agent newsroom — a supervisor handing out assignments, Claude writing, ChatGPT reviewing — and let it produce a daily briefing overnight; the outlet says the only human input was the original brief and one rejection the next morning. LobeHub's founder Wen Bin told the same outlet his team has not started a local dev server in three months and merges on the agent's acceptance report instead, which he puts at three to five times faster. The piece worth stealing is the plumbing: because Vercel caps a function at 300 seconds, each agent step is its own call with state in Redis, so a task can run for hours without the agent noticing. The team also open-sourced Acceptance, which turns a prompt into observable checks and runs them against the real product with screenshots and logs before a human approves. Expect setup friction — auth, credits, permissions — before it feels like a colleague.
Craft Anything (Rust, about 19,100 stars, MIT or Apache-2.0) does something agents are usually bad at: it studies a target before building. Point it at a website, desktop app, phone app or game, and the engine inspects the original with real tools, hands the findings to a coding agent that builds a new project in whatever stack you name, then compares the result against the original — pages at three screen widths, a game's own self-test — and refines until the gap closes. It ships 320 tools, each one both a command-line call and an MCP tool, so any agent can drive the same engine the desktop app uses. The design decision that matters is the default: clean-room mode refuses to read the original's code or files and works only from what a visitor can see, while Deep mode reads everything and can call Ghidra. That is a policy, not a sandbox — the README says so plainly — so it will not protect you from a coding agent acting with your permissions. Created on October 31, it has picked up 19,000 stars in a week and is still labelled early alpha.
unslop (JavaScript, about 11,000 stars) is a skill for Claude Code and Codex aimed at the specific look of vibecoded software: purple gradients, glows, emoji labels, shouty copy, uneven spacing. The skill names those tells first, sets a design direction from what the product actually is, then fixes the screen pass by pass the way a designer does a redline review. It works on Next.js, React, Vue, Svelte, Tailwind or plain HTML and CSS, on dashboards, apps and landing pages, and it can work from a live URL when all you have is a link. It also records before-and-after story videos, which is the part that makes it useful in a review thread rather than just in your editor. Eleven thousand stars in nine days says the AI-slop backlash has found its tooling.
Worth watching this week: whether LobeHub ships the agent-to-agent handoff it describes as its next level, and whether Craft Anything's clean-room promise survives contact with targets that fight back.
Which would you hand an agent first — your inbox, your codebase, or your design system? Write to us via our contact page.
