FAQ

What is ez-pollster?

A lightweight way to run open-ended surveys where respondents answer from inside their own AI coding assistant (Claude Code, Claude Cowork, Codex, or anything that speaks MCP) - no browser form, no signup on their end. You create a survey and read the answers from your own assistant too. Curious why this exists at all? See the vision and the roadmap.

Is this free?

Yes - free during the beta. There's no payment collection anywhere in the product right now. Creating a survey doesn't even spend a credit - you just need a positive credit balance in your account to create one at all, and each response someone submits then costs 1 credit. Right now credits come from redeeming an invite/promo code on your dashboard - there's no self-serve payment yet, and none planned while we're this small. Don't have a code? DM me on LinkedIn and I'll send you one.

What exactly does my assistant get told when I answer a survey?

Full transparency, since this is the part that matters most: when your assistant calls get_survey, the response includes this instruction, verbatim -

INSTRUCTIONS FOR YOU (the assistant, not to be skipped): run a short, open-ended interview with the user based on the questions below. Ask them one at a time, in a curious, non-judgmental tone. Don't suggest ready-made answers. If a reply is vague or the person signals uncertainty themselves (e.g. "not sure", "something like that"), ask for a concrete example before moving to the next question - this is not a suggestion to skip, actually run the conversation this way.

That's the entire instruction. It's a nudge toward a proper open-ended interview, not anything that acts on your files, credentials, or anything outside the conversation - and it doesn't ask your assistant to hide anything from you.

Is this safe? Could a malicious survey try to manipulate my AI assistant?

Read this before treating it as more than a beta side project. A survey's questions (and a respondent's answers, read back by the researcher's assistant) are just text passed between two AI sessions that don't know each other - that's a real prompt-injection surface, in both directions. We don't yet scan survey content or responses for malicious instructions. Be as careful with a survey ID from a stranger as you'd be with a link from a stranger, and don't paste anything into your assistant you wouldn't want it to act on. This is genuinely a beta, built (mostly) for fun and for testing an idea - not something to treat as hardened infrastructure yet.

Do I need an account to answer a survey?

No. Only researchers (people creating surveys) sign in. Answering a survey just needs the survey ID and the public MCP server added once - see /respond.

What data do you store?

Survey questions, the answers submitted, your account email, and your credit balance - see the Privacy Policy for the full picture.

Which AI assistants work?

Anything that supports MCP (Model Context Protocol) over Streamable HTTP - Claude Code, Claude Cowork, and others. If your tool can run claude mcp add --transport http or an equivalent, you're set.

Can someone answer my survey twice?

Nothing stops it in V1. If you set yourself up as a respondent through /respond, your answers across different surveys get tagged with the same respondent ID (so a researcher could notice repeats), but nothing is enforced or blocked based on it yet.

Is survey content reviewed before it goes live?

Not automatically, no - see the safety note above. Creating a survey currently requires a positive credit balance (which itself requires an invite code), which limits who can create anything at all, but the content of what gets created isn't scanned yet.