Talk to it
No CLI to learn.
After one init, you just speak to your agent — it translates your words into the deterministic steps.
![Real session: distill karpathy/micrograd, wear it, ask a question — answered in its idiom with [src:] citations](https://raw.githubusercontent.com/TTigger/mask/main/assets/demo.gif)
Distill anything. Wear anyone.
An agent-native, open-source framework that distills any source — a blog, a YouTube channel, a GitHub repo, a PDF — into a switchable persona the AI agent you already use can wear. Fully local, no API key.
The idea
mask is the deterministic toolbox; the intelligence is borrowed from the agent you already pay for.
The CLI calls no LLM. Your own agent (Claude Code, Codex, Cursor, Gemini) does the extraction by following a plain-Markdown recipe. Zero API key.
Every mask is a folder of Markdown + Git on your machine — human-readable and hand-editable. Your masks never leave your disk.
Every claim is cited to a source sample [src:…] and traceable to its origin. Thin evidence is declared, not hidden.
Talk to it
After one init, you just speak to your agent — it translates your words into the deterministic steps.
Quickstart
Node ≥ 20 — no Bun required. Your masks live separately in ~/.mask/.
One command from npm. Prefer to try first? npx mask-cli init works too.
npm i -g mask-cli # or: bun add -g mask-cliDefault installs the Claude Code orchestrator. For every other tool, write the universal AGENTS.md into your project with --agent agents-md --out .
mask init
No more commands — talk to your agent in natural language.
Per-source tools (only for that kind): git for repos, yt-dlp for YouTube, pdftotext for PDFs. Blogs need none.
Portable by design
Two adapters cover them all — Claude Code subagents, and the universal AGENTS.md standard that 30+ tools read natively. Add a source via an ingest module; the recipe stays put.
Plus re-distillation (update only what changed), a compounding knowledge wiki, opt-in headless scale mode for huge corpora, and coverage + wiki-integrity reporting. Read the full docs →
See it in action
Distilled end-to-end with mask's own recipes — every claim evidence-bound ([src:…]), each passing mask coverage with zero orphan pages, broken [[links]], or uncited claims. Install one in seconds:
mask try hung-yi-lee mask wear hung-yi-lee
Hung-yi Lee's Mandarin lectures on LLM internals — a white-board, example-driven teaching voice.
Karpathy's tiny scalar autograd engine — near-complete coverage of the house conventions.
Yiwei Ho's agent-native slide framework — arbitrary React pages on a fixed canvas, no DSL.
Chun-Yi Kuan & Hung-yi Lee's eval harnesses for two papers on hallucination in audio-language models.
See the gallery for the exact commands and sample answers.