Hi all. Disclosure first: we build Formgong, a hosted form backend, and this is our MCP server.
The problem we kept seeing: Cursor (or Lovable, v0, Bolt) generates a nice contact form, and it posts nowhere. Or it needs an API route + Resend + env vars that nobody set up. So we gave the agent a way to set up the backend part itself.
What it is
- Remote MCP server:
formgong.com/mcp(Streamable HTTP, stateless, JSON responses) - Auth: a personal API token,
Authorization: Bearer fgp_…, created in Dashboard → Account → API tokens. Tokens are scoped (forms:read,forms:write,submissions:read), can expire and can be revoked. initialize,pingandtools/listwork without a token, so you can look at the tools before signing up.
Tools
list_forms: your forms with id, name and public access keycreate_form: name + notify email, returns the access keyget_form_snippet: html, react or next code for one form, with the access key, the_langfield and the honeypot already filled inlist_recent_submissions: read-only and opt-in (separate scope); no IP or user agent; spam excluded by default
Cursor setup (~/.cursor/mcp.json or .cursor/mcp.json):
{
"mcpServers": {
"formgong": {
"url": "https://formgong.com/mcp",
"headers": { "Authorization": "Bearer ${env:FORMGONG_TOKEN}" }
}
}
}
Then something like: “Create a Formgong form called ‘Contact – acme.com’ and add it to the contact page as a React component.”
Where the submissions go: email and Telegram (Telegram is on the free plan too), plus signed JSON webhooks if you want n8n / Make / Zapier. Data is stored in the EU.
Limits, honestly
- No OAuth yet. Clients that only accept OAuth connectors need the
mcp-remotebridge. create_formis limited to 10 new forms per hour. Each token gets 60 requests a minute.- It doesn’t build forms visually. It creates the backend and gives the agent code to paste.
Free plan: 300 submissions a month. Paid plans start at $5/month (as of Oct 2026).
Docs (the Cursor-specific prompt is linked from there too): Formgong MCP server for Cursor, Claude and VS Code
I’d really like feedback on the tool design: are four tools the right cut, and what would you want the agent to be able to do that it can’t?