Open Source Radar — November 18: Cursor's 151-Plugin Marketplace

Today's board is about packaging: what an agent knows, shipped as a folder you install. Cursor's plugin repo is on trending again with 151 plugins, and two of the biggest skill libraries on GitHub are selling process rather than prompts.
cursor/plugins (TypeScript, 17,042 stars) is Cursor's official plugin repository and, more usefully, the reference for how the format works. Its marketplace manifest lists 151 plugins, 134 of them maintained by companies other than Cursor — Gmail, Salesforce, Figma, Shopify, GitHub, Playwright — which tells you who Cursor thinks will install these: teams wiring the editor into business tools, not one developer tuning a prompt. Two formats sit side by side. Cursor Plugins carry rules, agents, commands, hooks and variables on top of their own manifest; the open Agent Plugins format keeps just skills and MCP servers, and Cursor's docs say a spec-conformant plugin loads without changes. If you maintain an internal tool, the portable format is the one to target — the Cursor-specific parts are cheap to rewrite per editor, the encoded knowledge is not.
Claude-BugHunter (10,348 stars, MIT) is a skill bundle for bug hunting and red-team work — 83 skills and 15 slash commands built from 681 disclosed HackerOne reports across 24 vulnerability classes, with 433 of those reports individually cited so you can check the pattern against the original writeup. It splits into four layers: methodology, 58 web-app hunting skills, enterprise perimeter chains (M365, Okta, vCenter, SSL-VPN appliances), and a triage-and-reporting stage with severity mapping and evidence redaction. It hooks into Burp through MCP and the author says it has been run against DVWA, Juice Shop and Hacker101. The honest limit is the one every security agent has: it hands a tester the checklist a senior researcher already carries in their head, and scope discipline is a line in a markdown file, not a guarantee.
superpowers (319,683 stars, MIT) is a development methodology rather than a knowledge pack. It changes the first move: instead of writing code, the agent interviews you for a spec, shows it back in readable chunks, then produces a plan strict enough for a junior to follow — true red/green TDD, YAGNI, DRY — and only then fans the work out to subagents that review each other's output. The install list is the interesting part: Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, Qwen Code, Kimi Code, OpenCode, Devin CLI, Grok Build CLI and Hermes Agent all get a variant. Star counts in the hundreds of thousands are a poor quality signal on their own, but this one is at least selling the part most agent setups lack — a process the model cannot skip when it gets excited.
n8n (TypeScript, 210,833 stars) is the fair-code workflow platform that keeps landing on GitHub's daily board, and it shipped 2.48.5 this morning with two fixes aimed squarely at its AI Agent node: tool-call history now stays separate per input item, and streamed text no longer picks up stray newlines between chunks. Both are the kind of bug you only hit in production, when one workflow runs several items through the same agent and the outputs bleed into each other. It is not open source in the OSI sense — the license restricts resale and hosted competing services — so self-hosting is fine, embedding it in a product you sell is where you read the terms first.
Worth watching this week: whether the portable half of these bundles — skills plus MCP servers — becomes the thing everyone ships, leaving editor-specific rules and commands as local glue.
Would you install a stranger's 83-skill security bundle into an agent with shell access, or does that cross a line your team hasn't written down yet? Write to us via our contact page.







