Skip to content

Connect your agent

Server URL

https://mcp.redcrayon.ai/mcp

Every client signs in once, in the browser. Your agent then shows on the board as agent of <your name>, and the boards it draws are yours.

The plugin adds the server, the /draw:navigate command and code refs. Inside Claude Code you can also run /plugin marketplace add redcrayonai/draw, then /plugin install draw@redcrayonai.

Terminal window
claude plugin marketplace add redcrayonai/draw && claude plugin install draw@redcrayonai
  1. Start Claude Code in your project, run /mcp, choose draw, then Authenticate. Approve in the browser tab that opens.
  2. Run /draw:navigate. The agent opens a board named after your branch and posts its link. /draw:navigate <board link> draws on a board you already have.

Code refs and checkpoints need git and Node 20 or later on your PATH. Without Node, drawing still works.

Add the server to ~/.cursor/mcp.json:

{
"mcpServers": {
"draw": { "url": "https://mcp.redcrayon.ai/mcp" }
}
}

In Cursor Settings → MCP, click the login button next to draw and approve in the browser.

Terminal window
codex mcp add draw --url https://mcp.redcrayon.ai/mcp
codex mcp login draw

Customize → Connectors → Add custom connector. Name it draw, paste the server URL, then click Connect. claude.ai draws from the conversation; code refs need a coding agent in your project.

Any MCP client with remote servers and sign-in works: give it the server URL. VS Code uses "servers" and "type": "http".

If you connected before the rename, your redcrayon connection still works. To rename it to draw, remove that server entry in your client, add it again with the name above, and sign in again.

Outside Claude Code, start with:

Open a redcrayon board for this task and draw what you build on it as you go.

When a tool needs a capture of your checkout, it tells the agent the exact command, and the agent runs it. Nothing to install; Node 20+ and git on your PATH. To check your setup:

Terminal window
npx -y @redcrayonai/draw@1 doctor

Code snippet storage is off by default. A code ref keeps paths, symbols, line numbers and hashes without code text. For pushed GitHub code, readers can open the source with their own repository access.

Code refs and checkpoints still upload each cited file temporarily so the server can check it. The server reads it in memory. Captures expire within ten minutes and cannot be read again after the call claims them.

To enable inline snippets, open dashboard Account settings → Code → Store code snippets on my boards. The next ref or checkpoint can then store code text. Disabling this setting deletes stored snippets after a warning, without deleting ref metadata.

Sharing is a separate choice. View links exclude Code and Threats by default. Edit links always include code refs, with snippet text only if the owner consents to storage.

Your data and code consent explains uploads, operator access, deletion and backups. Hosted board data is not end-to-end encrypted.

Draw one request path through this repo. Cite each connection, post the board link, and listen for my red crayon marks.

Then review with the red crayon, investigate an incident, or propose an architecture change.

To point the Claude Code plugin at a server you run (Self-host quickstart), start Claude Code with:

Terminal window
REDCRAYON_MCP_URL=http://localhost:8788/mcp claude

Other clients take that URL in place of the hosted one.