Doan

Code shouldn't be your first draft.

Doan is a planning canvas for Claude Code and Codex. You think out loud in the browser; the coding session you already have open turns it into screens, flows, data and the questions you skipped. When nothing is left undecided, the plan goes back to that same session as a spec, and it starts building.

한국어

• Your session, your subscription. Doan runs on your machine and attaches to Claude Code as a channel, or to Codex as an MCP server. No API keys, no account with us, and no one else's login in between.

• It asks before it guesses. Every hole in the idea becomes a question on the canvas: structural ones (who approves? can one date go to several people?) block the handoff until you answer, defer or drop them.

• Colors mean something. Graphite is what you decided. Blue pencil is Claude's: new or proposed. Red is still open.

• The plan stays in your repo, in .doan/, next to the code it becomes.

Requirements

• One coding agent:

  • Claude Code 2.1.80 or later, signed in with a claude.ai account (Channels, research preview). Doan starts it with a development-channel flag, because custom channels aren't on the preview allowlist.

  • or Codex CLI, signed in.

• Bun 1.3 or later, and git.

• macOS or Linux. On Windows, use WSL.

Related MCP server: polyplan-mcp

Install

curl -fsSL https://getdoan.com/install.sh | bash

It clones Doan into ~/.doan, links doan into ~/.local/bin, and tells you which agents it found (read it first). By hand:

git clone https://github.com/Streamize-llc/doan ~/.doan
cd ~/.doan && bun install
ln -s ~/.doan/bin/doan ~/.local/bin/doan    # any folder on your PATH

Run the installer again to update.

Use

In the project you're planning:

cd ~/code/my-app
doan                        # with Claude Code; other claude flags pass through: doan --model opus
doan codex                  # with Codex; codex flags pass through: doan codex -m gpt-5

Nothing is added to your Claude Code or Codex settings: doan passes the MCP server for this session only.

1. Claude Code warns about the development channel. Choose I am using this for local development. (The startup line server:doan · no MCP server configured with that name is cosmetic; events arrive.)

2. The canvas opens at http://localhost:7420. A tab left open from last time reconnects on its own.

3. Write on the left, the way you'd explain it to a friend. When you pause, it's sent (untick Send when I pause to send by hand).

4. Claude answers in two steps: first the structure (screens, elements, actions), then critiques and questions.

5. Answer the open questions on the right (option chips answer in one click), and accept or dismiss proposals (dashed blue) and critiques.

6. When nothing structural is open, Send to Claude Code: the spec is written to .doan/specs/<id>.md, and the same session reports an implementation plan in the terminal and waits for your go.

With Codex, the same canvas and the same steps, with one difference: an MCP server can't push into a Codex session, so Codex waits on the canvas with a doan_wait tool and goes back to waiting after each change. While it waits, you plan in the browser, not in the terminal. If you interrupt it, the canvas says Codex isn't listening; tell it to continue. Doan's tools are approved up front for that session, so drawing doesn't ask for permission each time. After the handoff Codex stops waiting and shows its implementation plan in the terminal, like Claude.

Planning turns want speed more than depth, so doan starts Claude Code with --effort medium (about 18 s to the first structure, 33 s to questions and critiques; xhigh takes 80 s or more). Raise it with /effort high before building from the spec, or start with doan --effort high.

What Claude can draw

• Screens: full screens, modals, sheets and toasts (with the screen they sit on), named states ("Empty", "Pending approval"), screens only some roles see.

• Elements: header, text, button, input, list, card, image, grid, tabs, toggle, sheet, nav, divider, chat, picker, media, chart, map, progress, avatar, and one level of section / row grouping. Elements can be bound to data (T1.date).

• Systems: what happens off-screen: AI, background jobs, notifications, payments, third-party services.

• Data: fields, relations, retention, and whether it's sensitive (the Data tab).

• Actions: moves between screens and systems, conditional branches, role limits.

• Flows group screens once there are many; scope marks what ships first and what waits.

Canvas: drag to move, scroll to pan, ⌘/Ctrl + scroll (or pinch) to zoom, F to fit, double-click a screen name to rename it.

Where plans live

<project>/.doan/plans/<id>.json   # the canvas
<project>/.doan/specs/<id>.md     # the spec written at handoff

Try it without Claude

bun run demo    # draws a sample plan (shift swaps for nurses) with a scripted stand-in

Settings

Doan Team

Planning alone is free and stays free. Doan Team (getdoan.com) is our hosted service for teams who want to see each other's plans: every plan from a linked project syncs to your team, teammates open it in the browser (read only), and their comments land in the planner's chat as they're written and reach Claude on the next turn. Free for up to 3 members; larger teams subscribe.

doan team join <team-url> <token>   # the whole command is on the invite page; run it in the project folder

Teammates who don't plan can still bring the team's plans into their own agent: Doan Team is a remote MCP server at https://getdoan.com/mcp (sign in once in the browser), for Claude Code, Codex, Cursor or claude.ai.

The team service never calls Claude and never sees anyone's Claude login: every member plans with Claude Code on their own subscription, and only the plan (without the chat) is shared. This repository contains the client side (packages/team); the team server itself isn't open source and isn't offered for self-hosting.

License

MIT. See CONTRIBUTING.md to help.

Doan is not affiliated with or endorsed by Anthropic. Claude and Claude Code are trademarks of Anthropic.