Model Context Protocol
MCP is the interface an AI client uses to reach someone else's tools — ChatGPT, Claude, Cursor or an agent you built yourself. asksteps runs such a server. Four tools are open without signing in: an agent can check a PDF and draft a form before anyone has an account. For anything that touches your account it needs your permission.
Server address: mcp.asksteps.co
Four tools need no sign-in. The agent reads the plans, searches the public pages, checks a PDF for fillable fields — the file is not stored — and turns a conversation into a draft. What comes back is a link you can open without an account.
After you sign in through the browser, the agent can list your forms, read their statistics and fetch submissions. You grant each of those permissions separately.
Create a form, replace an existing one, store a PDF as a template. What the agent cannot do is publish — for that it gets a link so you can do it yourself.
The access belongs to plans from Starter up. On top of that you flip a switch in your account that is off by default. A plan alone unlocks nothing.
Form Builder
Drag elements from the palette onto the canvas, arrange them in rows and columns, spread them across several steps. Required fields are checked per step. Anyone who forgets something finds out right away, not after submitting.
Inbox
Submissions with a processing status and a note, free-text search, uploaded attachments, and the completed PDF to download. For spreadsheets, there's a CSV export.
Read forms, change forms, read submissions, PDF templates. "Read submissions" is deliberately not part of "read forms": a submission contains text other people wrote. Whoever reads it should not be allowed to change things in the same breath.
An agent may create and change. Putting a page on the open internet is a different decision, with legal obligations attached and an address that is in circulation afterwards. For that it gets a confirmation link, nothing more.
Every access you granted is listed in your account. Take one back and it stops in that moment, not when some token happens to expire.
An agent cannot change a form that came from a PDF: its fields are wired to boxes in the original document, and rebuilding it would break that wiring. It gets a refusal with the reason.
It works with a flat description of steps and fields, not with the internal configuration. If your form contains something that description cannot express, the write is refused — rather than quietly dropping it.
The connection is made over OAuth: you confirm which client gets which permissions. An API key is not accepted here — a key ends up in somebody's backup, a granted access does not.
Model Context Protocol. An interface an AI client uses to reach tools from other providers — roughly what the studio is for a person, but for an agent.
Any client that speaks MCP over HTTP, for example ChatGPT, Claude or Cursor. You enter the server address and the client walks you through signing in.
On the tier without an account it is not stored. It lives in memory for the length of one request; what comes back is the number of fields, the number of pages and the field types. Only with an account is a template kept — because then there is someone responsible for it.
Once you switch the access on and grant a client the "read submissions" permission: yes, to that client's provider. That is why the switch is off by default, why this permission is separate from the others, and why you can take it back at any time.
No. It is included from the Starter plan up.
The first tier needs no account: an agent can check a PDF and draft a form before you sign up.
Create an account