Skill Seekers turns documentation sites, GitHub repositories, local codebases, PDFs, Word and PowerPoint files, EPUBs, Jupyter notebooks, OpenAPI specs, AsciiDoc, HTML, RSS feeds, man pages, videos, Confluence and Notion wikis, and Slack or Discord exports (18 source types) into structured knowledge assets. One preparation pass exports to 22 targets: agent skills for Claude, Gemini, OpenAI, MiniMax, OpenCode, Kimi, DeepSeek, Qwen, OpenRouter, Together, Fireworks, or plain Markdown, documents and nodes for LangChain, LlamaIndex, and Haystack, upsert-ready data for Chroma, FAISS, Weaviate, Qdrant, and Pinecone, and Atlas and IBM Bob formats.
The pipeline scrapes a source (checking for llms.txt first, then a sitemap, then headless browser rendering for JavaScript sites), categorizes pages into topics such as API, guides, and tutorials, enhances the result with an AI-written SKILL.md and reference files, packages a platform-ready artifact, and optionally uploads it. Codebase analysis runs three streams (AST-based code analysis with GoF pattern detection across nine languages, documentation, and community issues and PRs) and can write how-to guides and architecture overviews. A project scan reads manifests, CI files, and sampled imports to emit one config per detected framework, and a community registry holds shared configs.
AI enhancement runs through one transport in two modes: API mode with Anthropic, Google Gemini, OpenAI, Moonshot/Kimi, or MiniMax, or local mode that drives an installed coding agent (Claude Code, Kimi Code, Codex, Copilot, OpenCode, or a custom command) without API costs. Enhancement depth is adjustable, and quality scoring can gate output below a threshold. Multi-source configs merge documentation, GitHub, PDFs, and videos into one asset with conflict detection across sources, and an opt-in SQLite FTS5 index lets an agent search references before reading them.
Finished skills install into 19 coding agents, including Claude Code, Cursor, VS Code with Copilot, Goose, OpenCode, Windsurf, Roo Code, Cline, Aider, Kilo Code, Continue, Kimi Code, and IBM Bob. A FastMCP server with 40 tools serves Claude Code, Cursor, Windsurf, VS Code with Cline, and IntelliJ IDEA over stdio or HTTP. Doc-change detection schedules re-scrapes, and streaming ingestion handles sets of 10,000 to 40,000+ pages.
Features
- 18 source types: documentation sites, GitHub repos, local codebases, PDF, DOCX, EPUB, Jupyter, OpenAPI, PPTX, AsciiDoc, HTML, RSS/Atom, man pages, YouTube/Vimeo/local video, Confluence, Notion, and Slack/Discord exports
- 22 export targets: skills for Claude, Gemini, OpenAI, MiniMax, OpenCode, Kimi, DeepSeek, Qwen, OpenRouter, Together, and Fireworks, documents for LangChain, LlamaIndex, and Haystack, vector data for Chroma, FAISS, Weaviate, Qdrant, and Pinecone, plus Atlas and IBM Bob
- Documentation scraping: three-layer discovery (sitemap, llms.txt, headless browser) for JavaScript sites, topic categorization, and a lenient HTML parser for broken markup
- Codebase analysis: AST parsing with 10 GoF pattern detectors across 9 languages, usage examples from tests, AI-written how-to guides, config extraction, and architecture overviews at quick, standard, or deep presets
- AI enhancement: API mode (Anthropic, Google Gemini, OpenAI, Moonshot/Kimi, MiniMax) or local mode through Claude Code, Kimi Code, Codex, Copilot, OpenCode, or a custom agent, with adjustable depth
- Project scan: an agent reads manifests, Dockerfile, CI, and sampled imports to emit one config per detected framework, with optional publishing to the community config registry
- Multi-source skills: one config merges docs, GitHub, PDFs, and videos into a single asset with conflict detection and pairwise synthesis
- Video extraction: subtitles, the YouTube transcript API, or local Whisper for transcripts, plus optional OCR of on-screen code from sampled frames
- Search index: an opt-in SQLite FTS5 index and stdlib search script inside the skill so an agent finds the right section before reading a large file
- Quality, sync, and scale: quality scoring with a threshold gate, doc-change detection with scheduled re-scrapes, streaming ingestion for very large doc sets, and incremental updates
- Agent installation: skills install into 19 coding agents, including Claude Code, Cursor, Copilot, Amp, Goose, OpenCode, Letta, Windsurf, Roo Code, Cline, Aider, Bolt, Kilo Code, Continue, Kimi Code, and IBM Bob
- MCP server: 40 FastMCP tools for Claude Code, Cursor, Windsurf, VS Code with Cline, and IntelliJ IDEA over stdio or HTTP, plus a GitHub Action and a Claude Code plugin in companion repos
