Get started with your AI assistant
Connect Claude, ChatGPT, Claude Code, Cursor or any MCP client to PanelWave in a minute, approve it once in your browser, and watch its first request land.
PanelWave has one page for connecting an AI assistant: Home → Connect an AI assistant (app.panelwave.org/connect). It shows the ways to connect, and underneath it a live box that follows the setup from the first request to the assistant's first tool call. Nothing on the page is a secret, so you can leave it open next to your assistant.
You get there from Home: a studio without a work shows How do you want to start? with the two paths (draw and letter in the editor, or let an AI assistant build it), and the Home checklist carries an optional step Connect an AI assistant that ticks itself once an assistant made its first call. The create-work wizard links here as well.
The assistant works as you, within the permissions you approve and your team roles. It asks before anything destructive or public. The artwork always comes from you; the assistant uploads and arranges, it never draws.
Coding agents: paste one line
For Claude Code, Cursor, Codex, Cowork and other agents that can run commands, copy the prompt from the page into your agent:
Set up PanelWave for me: run `npx -y @panelwave/mcp@latest setup --link j3k9qz` and follow what it prints.
The six-character code at the end is yours; it lets the page show your request. Without it (for example when you copy the line from this page) the setup still works, you just open the approval link the agent prints.
The agent runs the setup
It needs Node.js 20 or newer. The setup detects the clients on your computer (Claude Code, Claude Desktop, Cursor, VS Code, Codex, Gemini CLI), asks PanelWave for an approval link and prints it. An agent relays the link to you and waits; it never opens it itself.
You approve once in your browser
The link opens Approve your assistant in PanelWave (sign in if needed). You see which client asks from which computer, tick or untick the permissions (publishing and analytics are off unless you tick them), give the token a name and an expiry (a year by default), and click Approve. If you opened the page from the setup page, the request shows up there as well, with a Review & approve button.
The setup finishes on its own
Within a few seconds the agent's setup receives the token and writes it straight into the client's configuration: Claude Code through its own CLI (user scope), the others into their settings file, with a backup of the original. The token is never printed and never passes through the chat. The setup also writes a short brief for the agent into AGENTS.md in your folder and runs one check call.
Restart and ask
Restart the client (Claude Code: run /mcp or start it again). The agent then asks what you have already (a script, artwork, a work) and which start you want: a script into a work, a folder of panels into a work, or an existing work to translate. The page shows the first call as soon as it lands.
Prefer the terminal? Run npx -y @panelwave/mcp@latest setup yourself in the folder with your scripts and artwork; --help lists the options. Useful later:
| Command | What it does |
|---|---|
npx -y @panelwave/mcp setup --renew | Replaces the token (after it expired or was revoked) |
npx -y @panelwave/mcp setup --renew --scopes publish | Widens the permissions, for example to let the assistant publish review links |
npx -y @panelwave/mcp setup --client cursor --dirs ~/comics | Configures one client and names the folders the local file tools may read |
npx -y @panelwave/mcp setup --scaffold | Creates a project skeleton (concept.md, script.fountain, panels/, references/, panelwave-build.md) |
npx -y @panelwave/mcp status | Shows what is configured and whether the token still works |
npx -y @panelwave/mcp doctor | Checks Node, npx, clients, folders and the server, with a fix for each failure |
Claude and ChatGPT
Hosted assistants cannot run commands. Add PanelWave as a connector and sign in once; the page shows the address and the steps. See Connect Claude and Connect ChatGPT. To let Claude Desktop read scripts and upload artwork from your computer, install the Claude Desktop extension from the Claude tab: download panelwave.mcpb, open it, paste an access token and pick your comics folder.
The live box
| State | What you see |
|---|---|
| Listening | The page checks for a few minutes whether anything arrived; it pauses while the tab is hidden |
| Approval needed | Claude Code on JENS-PC wants to connect with Review & approve, when the request carries your page's code |
| Set up | The setup's own check call landed; the page now waits for the assistant's first request |
| Connected | Which app called which tool, how long ago, and a hint when the last call failed (for example a missing permission) |
Activity on your profile page lists the recent calls of every connected app: the app, the tool and the outcome, never the text you sent.
What to ask first
The page offers three starter requests with copy buttons:
- Turn
script.fountaininto a work with mobile and A4 pages; show me the outline dry run and create chapter 1 first. - Upload the images in the panels folder into chapter 1 and attach each one to its panel by file name.
- Add German, translate the balloons for review, and show me chapter 1 on a phone in German.
If you only have an idea, open No story yet? and ask the assistant to help with the concept and the script first. The assistant offers this path only when you ask for it.