MCP form backend AI agents speak natively.

It's a Model Context Protocol server that exposes splitforms as a typed toolset to AI coding agents. Instead of pasting your access key into Cursor or Claude Code and asking it to remember API docs, the agent calls structured tools like list_forms, get_submission, and generate_form directly. The server is free on every plan including Free.

Any MCP-compatible host: Claude Code (Anthropic's CLI), Cursor, Windsurf, Cline, Continue, and Zed are all officially supported and tested. Any other client that implements the Model Context Protocol spec works automatically — that's the whole point of the standard.

Free on every tier including the Free plan. We don't charge for AI access — submissions are still counted against your plan's monthly limit, but read/list/generate calls through MCP are unlimited and uncharged. The Pro and 3-Year plans subsidize the Free tier; AI access is a feature, not an upsell.

Sign in at splitforms.com, open Dashboard → MCP, and click 'Generate token'. The token starts with lim_ and is shown once. Copy it into your IDE's MCP config or use the Claude Code HTTP command on this page. Tokens are scoped to your account and can be revoked from the dashboard.

It's read and write. The MCP can list forms, read submissions, generate HTML, and call create_form to provision a new form on your account. All writes are scoped to the token's account — the agent cannot touch other accounts, billing, or your password. Destructive operations (delete) are intentionally not exposed.

Yes — MCP is host-agnostic. If you run a local model through Ollama, LM Studio, or vLLM and pair it with an MCP-aware harness (Claude Code, Continue, Cline all support local models), the splitforms MCP works identically. No data ever leaves the path between your IDE and our API.

The REST API is what you call from a browser or backend with fetch(). MCP is what an AI agent calls — it's a structured, typed tool layer on top of the same backend, with introspectable schemas the LLM uses to plan multi-step actions. You can use either or both. MCP exists so agents don't have to scrape your docs.

Yes, instantly. Open Dashboard → MCP, click 'Revoke' next to the token, and every future request from that token returns 401. Token usage is logged with timestamp, IP, and the tool called, so you can audit exactly what an agent did before revocation. Generate a fresh token whenever you want.

Twelve typed tools, all scoped to your token's account: list_forms and get_form read your forms and their schema; list_submissions and get_submission read submission data; generate_form and list_templates return ready-to-paste code from templates; get_access_key wires the public key into a generated form; create_form provisions a new form; update_form changes safe settings (name, subject, auto-responder, pause/resume); get_usage reports quota and storage; and create_webhook plus test_webhook wire submissions into Slack, a CRM, or Make/Zapier. There are also guided prompts (build_form, triage_inbox, connect_webhook, add_form_to_site). Destructive operations like delete a form or submission are intentionally not exposed over MCP — use the dashboard.

Anything the twelve tools cover. For example: 'Generate a Tailwind contact form for a Next.js server action that posts to splitforms', 'List the last 20 submissions to form_xxx and group them by source URL', 'Create a new form called Newsletter with fields email and consent_marketing', or 'Show me which submissions were flagged as spam this week and summarize the patterns.' The agent calls the tools directly to answer.

Because AI agents are first-class users now — in 2026 a meaningful fraction of new contact forms are written end-to-end by an agent, and making them scrape docs and copy keys is the same mistake as ignoring mobile in 2010. Copy-pasting an access key into an IDE chat leaves it in conversation history and screen recordings; an MCP token stays in a config file the agent reads but never quotes back. And an agent that can call generate_form and verify it against a real schema ships working code on the first try — docs go stale, a typed toolset can't.