# Librarian > Use for code research that needs dependency internals, upstream implementation examples, or external prior art. Always delegate to a subagent that investigates with opensrc and web search, then return only distilled findings, versions, paths, and links. - Skill: `sebastiaanwouters/librarian` (Agent Skill) - Install (CLI): `npx skillmds@latest add sebastiaanwouters/librarian` - Raw SKILL.md: https://api.skillmd.com/api/skills/sebastiaanwouters/librarian/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: sebastiaanwouters (https://skillmd.com/u/sebastiaanwouters) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/sebastiaanwouters/librarian --- # Librarian Always spawn a subagent first. If subagents unavailable, say blocked; do not do Librarian work in main context. Main agent frames the question, constraints, repo/package names, then waits and integrates only the subagent summary. Subagent first derives exact package/library version from this project: lockfile, manifest, tool config, workspace config, or vendored source. Research that version, not latest, unless user explicitly asks latest. Subagent uses `npx opensrc ` when internals matter; read `opensrc/` + `opensrc/sources.json`; fetch matching source before guessing. Subagent uses web search for newer ideas, docs, examples, prior art, and “who already solved this”; prefer primary sources, cite links, note dates when recency matters. Use `opensrc` for how code actually works. Use web for how others use it, compare patterns, or for anything likely newer than local context. Return only: answer, key evidence, exact package/repo/version and where version came from, relevant file paths/symbols, and links. No long logs, no pasted source, no wandering notes. Good prompts: “how is X implemented internally?”, “find examples of Y with Z”, “trace this error into dependency code”, “connect our code to upstream behavior”.