Best MCP Servers for Building Forms with AI Agents (2026)

What "MCP server for forms" means

An MCP server for forms is an HTTP (or local stdio) endpoint that speaks the Model Context Protocol and exposes a form product's operations as tools with JSON schemas: create_form, list_submissions, create_webhook, and so on. The AI host — Claude Code, Cursor, Claude Desktop, Windsurf, ChatGPT — reads the schemas once and then calls tools like local functions, with your credentials living in a config file the model never quotes back.

The category splits in two. Form builders (Tally, Typeform, Jotform) give the agent a hosted form or embed to hand you. Form backends (splitforms) give the agent an access key and a POST endpoint so it can write the form markup itself into your React, Next.js, Astro, or plain HTML project. If an agent is already editing your codebase, the second kind is the one that fits — see AI agent form backend for the longer argument.

What an agent can actually do with one

With the splitforms server connected, an agent can call 12 tools; every one is scoped to the account that owns the token, and all of them work on the Free plan. The table is taken from the server's live tool catalog.

Four bundled prompts sit on top of the tools: build_form (describe a form, get it created and paste-ready), add_form_to_site (detect the framework and insert the form), triage_inbox (summarize and rank recent submissions), and connect_webhook (wire submissions to Slack, Discord, a CRM, or Make/Zapier/n8n, then test delivery). In Claude Code they appear as /mcp__splitforms__build_form and so on. The full tool reference is on the MCP page and in the docs.

The ranking: 9 MCP servers for forms

Ranked by how much of the form job the agent can finish without you leaving the editor. Every entry below splitforms was checked against the vendor's own site or the community repository on September 6, 2026; the kind label says whether the server is first-party, community-built, or a generic automation layer that happens to reach form tools.

When not to use splitforms: if you need a visual builder, conditional logic between questions, payments inside the form, or a hosted form page you never embed, pick Tally, Typeform, or Jotform. splitforms is for forms whose markup you or your agent own.

Checked and left out

Two candidates were searched and omitted because we could not find a first-party MCP server. Formspree publishes agent instructions at formspree.io/ai — the agent fetches a page and follows a claim-URL flow — which is useful but is not MCP. Fillout is reachable only through community and third-party servers (danielma-tic on GitHub, Zapier, Pipedream, Composio), so it is folded into entry #9 rather than ranked on its own. If either ships a native server, this list will be updated.

Setup: connect the splitforms MCP server

The server lives at https://splitforms.com/api/mcp and authenticates with Authorization: Bearer <mcp_token>. Generate the token at Dashboard → MCP (it starts with lim_ and is shown once); the dashboard renders the same snippets below with your token filled in. The token is a secret — unlike the public form access_key — so keep it in the MCP config, never in app code.

Claude Code

claude mcp add --scope user --transport http splitforms \
  "https://splitforms.com/api/mcp" --header "Authorization: Bearer YOUR_MCP_TOKEN"

Restart Claude Code and run /mcp to confirm splitforms is connected.

Cursor

Save as ~/.cursor/mcp.json (or ./.cursor/mcp.json per project), then restart Cursor and check Settings → MCP.

{
  "mcpServers": {
    "splitforms": {
      "url": "https://splitforms.com/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_MCP_TOKEN"
      }
    }
  }
}

Windsurf

Save as ~/.codeium/windsurf/mcp_config.json. Note the key is serverUrl, not url.

{
  "mcpServers": {
    "splitforms": {
      "serverUrl": "https://splitforms.com/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_MCP_TOKEN"
      }
    }
  }
}

Claude Desktop

Claude Desktop's claude_desktop_config.json launches local stdio servers, so bridge to the HTTP endpoint with the community mcp-remote package. The dashboard does not ship a Claude Desktop tab yet; this is the standard pattern for any remote server that uses a bearer header.

{
  "mcpServers": {
    "splitforms": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "https://splitforms.com/api/mcp",
        "--header", "Authorization: Bearer ${SPLITFORMS_MCP_TOKEN}"
      ],
      "env": { "SPLITFORMS_MCP_TOKEN": "YOUR_MCP_TOKEN" }
    }
  }
}

Continue.dev is also documented in the dashboard. Any other client that supports streamable HTTP with custom headers works the same way. The endpoint is a single JSON-RPC 2.0 route: initialize, tools/list, tools/call, prompts/list, prompts/get, ping; batches of up to 25 requests are accepted.

Worked example: build_form → create_form → create_webhook → list_submissions

The whole flow is four tool calls, and the agent runs it from one prompt. Start with the bundled prompt (in Claude Code it is a slash command; in Cursor it appears in the prompt picker):

/mcp__splitforms__build_form
description: "quote request form for a landscaping company: name, email,
phone, service (lawn care / hardscaping / tree removal), project details"

The prompt compiles to instructions telling the agent to call list_templates first and use generate_form if a template matches, otherwise create_form with the exact fields. A quote form with a select does not match a template, so the agent sends this tools/call:

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "create_form",
    "arguments": {
      "name": "Quote request",
      "fields": [
        { "label": "Full name", "type": "text", "required": true },
        { "label": "Email", "type": "email", "required": true },
        { "label": "Phone", "type": "tel" },
        { "label": "Service", "type": "select",
          "options": ["Lawn care", "Hardscaping", "Tree removal"], "required": true },
        { "label": "Project details", "type": "textarea", "required": true }
      ]
    }
  }
}

name on each field is optional — the server derives full_name from "Full name" — and type defaults to text. The server inserts the form, then returns a text result the agent can paste straight into your project:

Created form "Quote request" with 5 fields — it's live in the splitforms.com dashboard.

access_key: 1f3c9e2a-…

WordPress shortcode (with the splitforms plugin):
[splitforms access_key="1f3c9e2a-…"]

Paste-anywhere HTML:
<form method="POST" action="https://splitforms.com/api/submit">
  <input type="hidden" name="access_key" value="1f3c9e2a-…">
  <input type="checkbox" name="botcheck" style="display:none" tabindex="-1">
  <label for="full_name">Full name *</label>
  <input id="full_name" type="text" name="full_name" required>
  <label for="email">Email *</label>
  <input id="email" type="email" name="email" required>
  <label for="phone">Phone</label>
  <input id="phone" type="tel" name="phone">
  <label for="service">Service *</label>
  <select id="service" name="service" required><option value="">Choose…</option><option>Lawn care</option><option>Hardscaping</option><option>Tree removal</option></select>
  <label for="project_details">Project details *</label>
  <textarea id="project_details" name="project_details" rows="4" required></textarea>
  <button type="submit">Send</button>
</form>

The botcheck checkbox is the honeypot the submit endpoint checks on every plan; the agent will usually adapt the markup to your stack (a fetch() handler in React or Next.js) while keeping the field names. Next, route submissions somewhere. create_webhook needs a public HTTPS URL and optionally a form_id from list_forms; Slack and Discord incoming-webhook URLs are auto-detected and formatted for those apps:

// create_webhook — scoped to the new form
{ "name": "create_webhook",
  "arguments": { "url": "https://hooks.slack.com/services/T000/B000/xxxx",
                 "label": "Slack #leads", "form_id": "<id from list_forms>" } }

// → result
{ "ok": true,
  "webhook": { "id": "wh_…", "label": "Slack #leads", "flavor": "slack",
               "scope": { "type": "form", "form_id": "…" }, "is_active": true } }

// test_webhook
{ "name": "test_webhook", "arguments": { "id": "wh_…" } }
// → { "ok": true, "webhook_id": "wh_…", "status": 200 }

Finally, after a real visitor submits, the agent (or the triage_inbox prompt) reads the inbox:

{ "name": "list_submissions", "arguments": { "limit": 5 } }

// → result
Showing 2 most recent submissions:

[
  {
    "id": "6b1e…",
    "when": "2026-09-06T14:02:11.000Z",
    "unread": true,
    "data": { "full_name": "Dana R.", "email": "[email protected]",
              "service": "Hardscaping", "project_details": "Patio, ~400 sq ft" },
    "from": "https://example-landscaping.com/quote"
  },
  { "id": "5a90…", "when": "2026-09-06T09:41:50.000Z", "unread": false, "data": { … }, "from": "Direct" }
]

get_submission with one of those ids adds the IP, user agent, and spam flag. Webhook payload details — including the files[] array with signed download URLs — are on the form-to-webhook page and in the API reference.

FAQ

Next steps

Have the agent add the form to a specific stack: Next.js, React, or Astro. Site built by an AI tool with a form that does nothing on submit? Read how to add a contact form to a site built with Claude, Cursor, Lovable, v0, or Bolt. Per-tool walkthroughs: Claude Code, Cursor. Not on an MCP host? The ChatGPT form backend and AI form generator cover the same endpoint without a protocol, and the contact form API page documents the POST itself.