boykush/scraps

The Wiki-link doc compiler for the LLM era.
Scraps treats documentation like a programming language. Wiki-linked markdown becomes a typed source, compiling into a static site for readers and into JSON any agent can shell into — turning Karpathy's *LLM Wiki* pattern into a typed, queryable artifact. CLI primary with companion AI skills, fitting any editor and any LLM agent.
Quick start
See the Getting Started tutorial for the full flow.
How it works
Scraps reads [[wiki-link]], #[[tag]], ![[embed]], [[Page#heading]], and ctx_path as typed primitives. The same source compiles to an HTML site for human readers and to JSON for scripts and AI agents. The IR is kept under .scraps/ next to .scraps.toml, so every command reads the wiki through it and an unchanged source is never parsed twice.
AI integration
CLI + JSON is the primary path — any shell-capable agent can query Scraps without an MCP client implementation. Bundled plugins provide agent-facing workflows:
- llm-wiki — Karpathy-style *Ingest / Query / Lint* skills for Claude Code and Codex, plus Claude Code agents for purpose-driven lint and the default Scraps LLM Wiki schema
- pkm — Personal knowledge management skills for Claude Code: *learn* a source and record only what you understood, or *digest* an article into a short linked page
- mcp-server — MCP server for MCP-compatible clients, shared across repositories via scraps mcp serve --http
See the AI integration guide for the trade-offs.
Documentation
- Documentation site — Tutorial, How-to, Reference, Explanation
- Sample wiki — Japanese knowledge base built with Scraps
Screenshots
*Search, pagination, and Wiki-link navigation, themed with Nord.*
*Each scrap page draws its neighbourhood: backlinks on the left, links on the right, arrows pointing at the referenced scrap.*
Light mode
Contributing
Bugs, feature requests, and PRs are welcome. See CONTRIBUTING.md, the issue templates, and the Code of Conduct.