Coding-Agent Form Backend: Give Claude Code & Cursor Real Data Collection

Claude Code, the Codex CLI and Cursor's agent mode write excellent frontend code and terrible infrastructure. Ask one to "add a contact form" and it will generate a beautiful form, wire up handleSubmit, then shrug: the POST has nowhere to go. A local dev server is not a backend, and an agent that invents api.example.com out of thin air is shipping a dead button.

Agents fail at collection for one reason: they can't read the docs you would have read for them. Fix that and the same agent will do the rest. Here is the two-part handshake—machine-readable documentation, plus a CLI the agent can run.

Handshake one: an endpoint that documents itself

Publish any form or assessment on FormLM and you get a share endpoint. Append ?help to it and the endpoint returns its own API documentation as Markdown—write format, query paging, summary aggregates, rate limits, field creation rules:

curl "https://formlm.me/api/v3/share/YOUR_TOKEN?help"

Drop that URL into the agent's context and it learns the whole surface in seconds. From there the rules are the ones any agent codes against: POST JSON as {"data":{...},"meta":{...}} (flat objects are accepted too), or submit a plain HTML form to /form; read records back from /query, aggregates from /summary. Unknown keys create their own fields, so an agent adding an input to the page never needs a migration.

Handshake two: one npx command builds the assessment itself

Collection is half the job—the other half is what makes a submission worth collecting: scoring logic, a per-respondent report, a published page. formlm-cli is the official MIT-licensed CLI on npm (Node.js ≥ 18), and because a coding agent already owns a terminal, it can drive the whole build:

npx @formlm/cli smart generate "Digital marketing maturity diagnostic for agency prospects, 5 dimensions, scored PDF report"
npx @formlm/cli snapshot --summary
npx @formlm/cli share url

The first line is the interesting one: smart generate turns a one-sentence brief into a complete assessment app—questions, weighted scales, conditional logic, report layout—cached server-side so the agent can execute it module by module. The second dumps the current state as JSON so the agent knows the field names it must submit against. The third returns the share token and page URL: the exact string that goes into action= or fetch() in the code the agent is writing.

A prompt that actually works

Agents follow contracts better than intentions. This is the shape of a request that ends with live data collection instead of a shrug:

<p>Before wiring the form, run: npx @formlm/cli smart generate "<your brief"
Then read the API contract at https://formlm.me/api/v3/share/<token>?help
Build the page's submit against that contract: POST {data, meta},
show an inline success state, and add a company honeypot hidden with CSS.
Verify with a curl round-trip before you call it done.</p>

Four things to notice: the agent creates the assessment rather than assuming one exists; it reads the contract rather than guessing field shapes; the honeypot is specified because generated pages never add spam guards on their own; and the curl round-trip gives the agent a test to pass—the same one a human would run.

What stays on you

• The token is a write target, not a credential. Share tokens authorize submissions to one form. They are safe in client code; Access Tokens (account control, 7-day expiry) never belong in a prompt or a repo.
• Spam still shows up. Rate limiting is on the endpoint; a honeypot field and read-time filtering are the generated page's job—ask for them explicitly.
• Scoring lives on the platform, not in the agent. That is the point: weights and report logic stay stable no matter how many pages the agent regenerates.