Drawa

A canvas workspace for Claude Code.

Run several Claude sessions side by side, watch every file they read, edit and run as a live map, and review every change without leaving the browser.

Install · Quick start · Features · Shortcuts · Contributing

Why Drawa

In a terminal, Claude Code's work scrolls past as a transcript. Drawa lays it out as a map instead. Each session is a card on a canvas. The files it touches appear in a window wired to that card, colored by what happened to them (read, edited, written). The commands it runs collect beside it. You can see at a glance what Claude did, and click any file to see exactly how it changed.

Drawa runs entirely on your machine. It drives the claude CLI you already have, with your login, settings, skills and MCP servers, and serves the UI on 127.0.0.1.

Features

• Parallel sessions. Each card is a live Claude process. Type any time; messages queue while Claude or its sub-agents work.
• Your work, drawn as a map. Files Claude reads, edits and writes are linked to its card and colored by action. Commands it runs collect in their own window.
• Diffs everywhere. Click a file for its changes from every session, plus the file itself with Markdown and Mermaid preview.
• Sub-agents in the open. Each sub-agent gets its own window showing its work as it happens, and you can message it directly.
• Claude can use the canvas too. Claude can read, create, edit and link canvas items through built-in tools, with changes going through the normal approval flow.
• A whiteboard around your code. Sticky notes, a Markdown scratchpad, Mermaid diagrams, pictures, pinned snippets, shapes, arrows and freehand ink. @-reference any of them in a message, or drop them on a card.
• Git and GitHub built in. Browse status and history. Open pull requests and issues through gh, and send a PR, its failing checks or its reviews straight to Claude.
• Picks up where you left off. Resume past sessions from history. The whole layout survives a reload, and a reload re-attaches to sessions that are still running.
• Themes. Light and dark, with Rosé Pine, Catppuccin, Tokyo Night, Gruvbox, Nord, Kanagawa, Everforest, Solarized and GitHub color schemes.

Requirements

No Claude subscription? Use OpenCode

OpenCode is an open-source coding agent that works with most model providers (OpenRouter, Google, OpenAI, local models and more). With it installed, New session ▾ in the toolbar offers an OpenCode session next to a Claude Code one, and each session card picks its model from OpenCode's list.

1. Install it: curl -fsSL https://opencode.ai/install | bash (or see its docs).
2. Connect a provider with opencode auth login, or skip that: OpenCode's own free models work without a key.
3. Start Drawa as usual. Its startup check lists opencode, and warns if your version isn't the one Drawa was tested with.

Things to know:

• Privacy. A session's prompts, and the files it reads, go to the provider you pick. OpenCode's free models are run by third parties that may use what you send to improve their models (check OpenCode Zen's terms), so don't use them on code you can't share.
• What's the same. Approvals, questions, the files and commands windows, sub-agents, canvas tools and history work as they do for Claude Code.
• What differs. OpenCode has no Auto mode, and no effort setting. The status line's usage windows are Claude Code's, so OpenCode cards don't show them.
• Memory. Each OpenCode session runs its own opencode server (about 300 MB). There's no OpenCode-specific limit: set DRAWA_MAX_LIVE if your machine needs one.
• Commit messages. The Git window's Write with ▾ picks which agent writes commit messages and pull request descriptions.

Use Codex

Codex is OpenAI's coding agent. With codex installed and logged in (codex login), New session ▾ offers a Codex session, and each card picks its model from Codex's list.

• What's the same. Approvals (commands, file edits, canvas changes), the files and commands windows, canvas tools, pasted images and history work as they do for Claude Code. Looking at the canvas never asks.
• What differs. Codex has no Auto mode and no effort setting here (it uses its default). Read only asks before edits and before any command Codex doesn't consider safe. Plan only never asks, nothing it runs can write, and it doesn't hand you a plan to approve: it answers in the chat. Allow edits approves file changes inside the project only (not in .git, .codex, .agents or .claude, and not renames out of it). Always allow covers a command or edit until the card's Codex process closes (after 30 idle minutes, or to stay under DRAWA_MAX_LIVE; it asks again after resuming), and Drawa doesn't save Codex's permanent command rules; canvas changes ask each time. A mode change applies from the next turn (a message sent mid-turn joins the running one), except that leaving Allow everything stops a running turn, since it can't lose full access mid-turn. Sub-agents show as one row, not their own window, and a queued message can't be taken back.
• Trust. For a local folder, Drawa marks it trusted for Codex, so, as with Claude Code, its .codex/ settings and AGENTS.md load. A cloned GitHub repo you didn't trust is marked untrusted, so they don't. Either way nothing is written to ~/.codex/config.toml.
• Memory. Each Codex session runs its own codex app-server (about 250 MB). There's no Codex-specific limit: set DRAWA_MAX_LIVE if your machine needs one.

Install

curl -fsSL https://raw.githubusercontent.com/HimalayanNomads/drawa/main/install.sh | sh

This downloads the right release binary for your OS and architecture and installs it to ~/.local/bin/drawa. You don't need Go or Node.

Update later with drawa --update. For manual downloads, custom install locations and uninstalling, see INSTALL.md.

Needs Go 1.22+ and Node 20.19+ (or 22.12+).

git clone https://github.com/HimalayanNomads/drawa.git
cd drawa/web && npm install && npm run build
cd .. && go build -o drawa .

Quick start

cd ~/my/project
drawa

Drawa opens http://127.0.0.1:8765 in your browser, and Claude works in that folder. Press N for a new session and start typing.

drawa                  # the current folder
drawa ~/some/project   # any other folder
drawa --net .          # also reachable from other devices on the network
drawa https://github.com/owner/repo   # clones it into a cache folder and opens it (reruns reuse the clone)
drawa --clean          # removes cached clones (or pass one repo's URL)

Opening a GitHub URL, or a folder inside its clone, asks on each launch whether to trust the repo (only the server's own restarts after code changes reuse the answer). Answer n (the default) and its own .claude/ settings, hooks, .mcp.json, CLAUDE.md, OpenCode project config, and Codex's .codex/ config and AGENTS.md are ignored, since they could run commands on your machine.

--clean keeps any clone with uncommitted, unpushed, stashed or ignored files. Don't run it while a Drawa is open on that clone. A GitHub folder link (/tree/main/docs) opens the repo on its default branch.

In the message box, / opens skills and slash commands and @ references files or canvas items. The toolbar picks the model and permission mode.

By default Drawa only answers on localhost. Pass --net (in any position) and, like a Vite or Next.js dev server, it also listens on the machine's network address and prints both:

  - Local:   http://127.0.0.1:8765
  - Network: http://192.168.1.23:8765/?token=fpKZrzCN

It also prints a QR code of the Network link, so a phone can open it with its camera. The Network link lets another device on the same network — a laptop, a phone — open the same workspace. It only works with its ?token= (a fresh one each run, checked once, then kept in a cookie and dropped from the address), and an address that guesses wrong 5 times locks out for a few minutes. See Security.

Configuration

DRAWA_PORT=8766 CLAUDE_CONFIG_DIR=$HOME/.claude-work drawa ~/other/project

Keyboard shortcuts

Shortcuts follow Excalidraw's where the tool exists, and don't fire while you're typing. Press ? in the app for this list, and Ctrl+K to run any action by name. A short tip shows in the corner at launch; Hide tips turns them off.

Mouse: drag the background to select (Select mode) or pan (Hand mode). The middle button and the wheel always pan. Ctrl/Cmd+scroll or pinch zooms; Shift+scroll scrolls sideways. The zoom is saved with the canvas. Jumping to a window (Ctrl+K, C, a notification) keeps the zoom unless the window would be under 50% or wouldn't fit, then it zooms to fit that window (at most 100%). Click a picture or diagram for full view, and click it again to zoom and pan it. Phones start in Hand mode.

Security

Drawa is built to run locally for one user:

• The server listens on 127.0.0.1 only, unless you pass --net (see Quick start), and checks the Host header on every request either way.
• With --net, the network address additionally needs its one-time ?token=; wrong guesses lock that address out after 5 tries. Still only pass --net on a network you trust, and treat the printed link like a password — don't post it anywhere public.
• Changing requests must come from Drawa's own page (a matching Origin).
• File access is confined to the project folder you opened.
• Claude's canvas tools use a per-process token, so only that session's own claude process can call them.

To report a vulnerability, see SECURITY.md: please open a private security advisory rather than a public issue.

Contributing

Issues and pull requests are welcome. Maintainers cut the releases; a merged pull request ships in the next one. See CONTRIBUTING.md for the development setup and an architecture overview, and CLAUDE.md for the conventions the codebase follows. Everyone taking part is expected to follow the Code of Conduct.

License

Drawa is licensed under the GNU Affero General Public License v3.0.