@agentic-survey/mcp-server

An open-source (MIT) MCP server that lets your own AI agent build surveys and read results, straight into your own Supabase. You own the data, you carry the AI cost, and your keys never leave your machine. It's free.

Part of agentic-survey-mcp. Site: mcpsurveys.com.

Why

• You own the data. Surveys, questions, and responses live in your Supabase project. Nothing is stored on anyone else's infrastructure.
• You carry the AI cost. Authoring and analysis happen inside your agent. There's no AI running server-side, so there's nothing to bill you for.
• Bring your own keys. Your Supabase secret key stays in a local config file. It's never transmitted, never logged, and never a tool argument the agent can read.

Quick start

# 1. Print the schema, paste it into your Supabase SQL editor
npx -y @agentic-survey/mcp-server init --print-sql

# 2. Store your keys locally and verify the connection
#    (typed into your terminal, never sent to the agent)
npx -y @agentic-survey/mcp-server init

Then point your MCP client at the server:

# Claude Code
claude mcp add agentic-survey -- npx -y @agentic-survey/mcp-server

// Claude Desktop / Cursor / other (claude_desktop_config.json, .cursor/mcp.json, ...)
{
  "mcpServers": {
    "agentic-survey": { "command": "npx", "args": ["-y", "@agentic-survey/mcp-server"] }
  }
}

Now just talk to your agent:

"Build a 4-question customer-satisfaction survey with a 1-5 rating, a yes/no, a multiple choice, and an open comment. Publish it and give me the share link."

"Summarise the results so far, themes from the comments too."

The agent chains the tools for you: create_survey → add_question → publish_survey → get_results.

Tip: npm i -g @agentic-survey/mcp-server gives you the shorter agentic-survey init command.

Do I need to enable RLS / anything in the Supabase dashboard?

No. The init --print-sql migration enables Row Level Security and the access policies for you, on all four tables. Don't disable RLS: it's what fences the publishable key (which travels in the share link) so it can only read published surveys and insert responses. It cannot read anyone's responses or your drafts. Your secret key, stored locally on your machine, bypasses RLS for authoring and analysis.

How it works

  your AI agent ──(MCP / stdio)──▶  mcp-server ──┐
                                                  ├──▶  your Supabase
  respondent ──(browser)──▶  page-service ────────┘     (source of truth)

Publishing a survey returns a share link carrying your project ref and publishable key. A respondent opens it, the stateless page-service writes their answers into your Supabase, and your agent reads them back with get_results. Supabase is the backend for now; more are planned.

Tools

setup_connection, create_survey, add_question, update_question, remove_question, reorder_questions, set_question_logic, validate_survey, publish_survey, get_share_link, list_surveys, get_survey, list_responses, get_results.

Question types: single_choice, multi_choice, rating, yes_no, number, short_text, long_text, date, time (24-hour), slider (0–100 percentage by default).

Branching / skip logic

set_question_logic makes a question conditional: it shows (or hides) based on the answers to earlier questions — e.g. only ask the follow-up if someone rated you ≤ 2. The page-service evaluates this live and submission validation respects it (a hidden required question isn't required; answers to hidden questions are rejected). Run validate_survey after wiring logic and before publish_survey — it catches forward/dangling references, conditions citing options that don't exist (so a question could never appear), and choice questions with no options.

License

MIT. Built by Monty. Source and issues on GitHub.