Skip to content

Claude Code & coding agents

Claude Code, Cursor, VS Code, Codex CLI and similar coding agents connect to the same remote server as Claude.ai and ChatGPT: https://mcp.zandovi.com/mcp. You sign in to Zandovi once in the browser (OAuth), approve the connection for one of your organizations, and the agent gets the 18 tools, including search_stock_images, that the scopes you approve allow, plus two prompts (design_template, fix_violations) that walk it through validate → preview → create. No API key and no local install are needed.

  1. Add the connector.

    Claude Code (CLI):

    claude mcp add --transport http zandovi https://mcp.zandovi.com/mcp

    Settings → API Keys → Connect an AI assistant shows this command with the right URL for your environment.

  2. Sign in. In Claude Code, run /mcp, pick zandovi and choose Authenticate. Your browser opens the Zandovi consent page: check the client name and the permissions it asks for, pick the organization, and approve. Claude Code stores the token and refreshes it on its own.

  3. Confirm the connection. Ask the agent something that uses a tool, e.g. “list my Zandovi projects” or “what’s my Zandovi account”. A working connection returns your organization, plan, quota and the connection’s scopes. Then try “design a business-card template for Acme and save it to my Cards project”: the agent takes the size from the presets, validates the design for free, previews it (one render) and saves it (saving renders once more), and gives you a link to open it in the designer.

Cursor reads MCP servers from mcp.json (Settings → MCP, or ~/.cursor/mcp.json for a global server). Add the URL; Cursor asks you to sign in the first time it connects:

{
"mcpServers": {
"zandovi": {
"url": "https://mcp.zandovi.com/mcp"
}
}
}

Add the server to .vscode/mcp.json (or run MCP: Add Server and choose HTTP):

{
"servers": {
"zandovi": {
"type": "http",
"url": "https://mcp.zandovi.com/mcp"
}
}
}
codex mcp add zandovi --url https://mcp.zandovi.com/mcp
codex mcp login zandovi

Any client that supports remote (Streamable HTTP) MCP servers with OAuth works the same way: add https://mcp.zandovi.com/mcp and sign in when asked. Zandovi does not ship a local (stdio) package; a client that can only run local commands, or a headless job that cannot open a browser, should call the REST API with an API key instead.

Coding agents get Zandovi’s design knowledge the same way as every other client, from one tool: get_design_reference({topic}) returns a whole knowledge-pack section: guide (design gotchas plus every element field’s default value), variables (binding variables, values and batch rules), codes (every validation code, split out to keep guide short), fonts, presets, examples, example:<name> for a full example design, elements for an index of all 15 element types, or element:<type> for one type’s field table, coordinate origin, unsupported fields and a valid example. Icon names come from search_icons. Both run locally against the bundled pack, need no API call and no credentials, and behave the same in every client. The same pack is also published as zandovi:// MCP resources, which an agent that supports MCP resources can read instead. The tools are the dependable path: when Claude Code reaches Zandovi through a connector added on claude.ai rather than with claude mcp add, the resources aren’t available.

Open Settings → Connected apps in Zandovi, click Revoke next to the agent and confirm. This invalidates its tokens immediately; remove the entry from the agent’s MCP configuration to stop it from trying to reconnect.