Drafty — Comentários Inline em Artefatos Claude
drafty-im/drafty
Feedback colaborativo para artefatos Claude — comente, Claude edita em tempo real
Ver no GitHubPlugin que permite compartilhar artefatos Claude como links para feedback em tempo real, estilo Figma. Você comenta em qualquer linha, Claude lê os comentários e edita o documento no mesmo link com histórico de versões. Funciona nativamente em Claude Code com um simples comando.
Instalação
npx convez-ai drafty-im/drafty1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Revisar especificações e planos com times em tempo real
- Iterar documentação e conteúdo sem copy-paste
- Coletar feedback paralelo em código e designs antes de versão final
Veredito

Ferramenta prática que elimina atrito no loop feedback-iteração. Integração nativa com Claude Code e interface sem configuração torna-a acessível, mas depende de estar online no drafty.im e da disposição do Claude em ler comentários — funciona bem para equipes pequenas com workflows ágeis.
README
A tradução para o português deste README ainda está na fila.
Drafty is point-and-comment review for the things Claude makes. Claude writes a plan, a spec, a page — you publish it to a link, then click any line and leave a note, Figma-style. Claude reads the comments and ships a new version on the same link, with history. No screenshots, no re-pasting "the third paragraph, the one about pricing."
Easiest: in a Claude Code session, just ask — "install the drafty-im/drafty plugin." Claude runs it for you (you approve the install once, or it's automatic on auto-approve), then run /reload-plugins to switch it on live — no restart.
Or do it by hand:
/plugin marketplace add drafty-im/drafty
/plugin install drafty@drafty-im
/reload-plugins
/reload-plugins activates everything live in the current session — the drafty skill and the drafty command on PATH, no restart. (Requires bun.)
How it works
Once installed, just tell Claude to "drafty it" after it writes something:
- Claude runs
drafty push <file>and hands you adrafty.im/canvas/<slug>link. - Open it. Hover any element, click, leave a comment. Share the link — anyone comments as a guest, no sign-up, live cursors.
- Tell Claude "address the canvas" (or set it live) — it reads each thread, edits the source, and pushes a new version on the same link. Old versions are kept.
You talk; Claude runs the commands. You never touch the CLI yourself.
What's in the box
- The
draftyskill — teaches Claude the whole loop: publish, read comments, reply on the canvas, mark threads done, push revisions, roll back. Claude loads it on its own when you say "drafty it" / "share this for feedback" / "what did they comment". - The
draftyCLI — a thin HTTP client (no keys — it starts as a persistent guest stored in~/.drafty, anddrafty loginopens your browser to sign in, one sign-in covering web + CLI, when you want to keep canvases).push,watch,inbox,reply,resolve,mode,claim, and the rest.
Modes
A canvas has one mode, and Claude sets it from how you talk:
| Mode | Viewers comment | Claude acts on comments |
|---|---|---|
readonly | no | — |
feedback (default) | yes | no — parked until you say go |
live | yes | yes — works them as they arrive |
"Go live" arms a realtime doorbell so Claude reacts the moment you comment; "park it" stops it.
Privacy & telemetry
The CLI starts as an anonymous guest — no account, no email — until you choose to drafty login (which opens your browser) to keep canvases under a real account. It sends basic usage events (e.g. canvas.published) to drafty.im so I can see what's used; set DRAFTY_NO_ANALYTICS=1 to turn that off. It only ever talks to drafty.im (override with DRAFTY_BASE_URL).
Links
- The product: drafty.im
- Agent quickstart (no install, demo only): drafty.im/get
MIT licensed.