How to Connect AI Agents to Real Consumer Research via MCP

An AI agent that can call a research server stops guessing what consumers think and starts asking them. This quickstart is the setup reference for that connection: why it is worth making, what you need, the exact setup for every client, a smaller tool profile if you want one, how to confirm it works, what to do if the tools don’t appear, and a first brief to send. Every client reaches the same User Intuition server at https://mcp.userintuition.ai/mcp, so once you are connected the research workflow is identical everywhere.

Why connect an AI agent to a research server?

An agent without a research connection answers questions about customers from its training data, which means it predicts what people might say instead of reporting what they said. Connecting it to a research server over the Model Context Protocol changes the source of the answer. The agent can brief a study, recruit real participants, and bring back findings with the quotes behind them, all from the conversation where the question came up. MCP makes that a setup task rather than an integration project: the client discovers the research tools when it connects, so nobody writes glue code or maps endpoints by hand. The connection is also portable, because the studies and reports belong to your account rather than to any one client. With User Intuition, the evidence comes from 20–30 minute interviews that ladder 5–7 levels deep, run by voice, chat or video with a vetted 4M+ consumer panel or your own customers.

Every panelist call is reviewed by the User Intuition team, panelists pass a video identity check, and only interviews that pass quality scoring are billed; see how quality works.

What do you need before you connect?

Three things, and only one of them takes more than a minute.

• A User Intuition account. Create a free account. Signup needs no card and includes 3 free interviews with your own participants; panel recruiting is billed separately.
• An MCP client. Claude, ChatGPT, Claude Code, Codex, Cursor, VS Code, or any agent framework with an MCP client.
• An API key, only for local stdio or REST. Create one under Settings, then API Keys, in the app. Hosted connections sign in with OAuth, so there is no key to store or rotate.

How do you connect each client?

Find your client in the table, then follow its section below. Hosted connections sign in with OAuth; the local stdio package and the REST API use an API key.

Claude (web and desktop)

Open Claude’s connector directory, find User Intuition Research, connect it and sign in. That connector alone runs the full research workflow, from the brief to the cited report. Because Claude supports MCP Apps, plan reviews and cited answers appear as interactive cards in the chat. The Claude directory guide lists example prompts, and running user interviews from Claude shows a complete run.

ChatGPT

Create a developer-mode connector with https://mcp.userintuition.ai/mcp, and ChatGPT discovers the OAuth sign-in on its own; the ChatGPT connector guide walks through each screen.

Claude Code

Choose one of two routes. The first installs the user-intuition-research plugin, which includes the hosted server and adds ten research skills:

/plugin marketplace add user-intuition/skills
/plugin install user-intuition-research@user-intuition

The second adds the server alone:

claude mcp add --transport http userintuition https://mcp.userintuition.ai/mcp

Pick one route rather than both, then run /mcp and sign in. The skills library describes each skill the plugin adds.

Codex

Add the hosted server, then sign in if the add command did not start the sign-in on its own:

codex mcp add userintuition --url https://mcp.userintuition.ai/mcp
codex mcp login userintuition

Start a new task so Codex loads the research tools.

Cursor

Add the hosted server URL in Cursor’s MCP settings and sign in when prompted. To run the server locally instead, add this entry to .cursor/mcp.json and restart Cursor:

{
  "mcpServers": {
    "userintuition": {
      "command": "npx",
      "args": ["-y", "@userintuition-ai/mcp@latest"],
      "env": {"USERINTUITION_API_KEY": "ui_sk_your_key_here"}
    }
  }
}

Read real keys from your shell environment so they stay out of version control. The @latest tag pulls the current package each time Cursor starts it.

VS Code and other stdio clients

Any client that starts a local MCP process runs the same package, npx -y @userintuition-ai/mcp@latest, with USERINTUITION_API_KEY in its environment. In VS Code, add that command and variable to the MCP servers setting and reload the window. Every key begins with ui_sk_, which makes a pasted value easy to check.

Custom agent frameworks

Agents built on LangChain MCP adapters, CrewAI or any framework with an MCP client launch the stdio package as a subprocess, with the key in the subprocess environment. A hosted service that cannot start a local process can connect to the hosted endpoint with an organization API key instead of OAuth. Either way, the agent discovers the same research tools that the chat clients see.

Scripts and CI

Install the command-line interface with npm install -g @userintuition-ai/mcp, then sign in through your browser with userintuition-mcp login. From the shell you can estimate panels, inspect fielding and read results, with JSON on stdout for scripts to parse, while a paid panel launch runs from an MCP client or the REST launch-panel endpoint. The CLI overview and CLI recipes cover the commands, and user research from Cursor and Codex shows the CLI inside an editor workflow.

Backend services over REST

Server-side code that already knows which call it needs can send an API key to the REST API without MCP. The developer integration guide walks a study through every request, and MCP vs REST for user research helps you decide which one a given job should use.

Can you connect with a smaller set of tools?

Every client can load a smaller profile instead of the full catalog. Point a hosted connection at the profile’s URL, or add the flag to the stdio command.

Each hosted URL signs in on its own, so authorize the one your client uses. The read-only profile suits an agent that should answer from finished studies without creating new ones, such as an assistant that drafts a weekly insights summary.

Which client should you connect first?

Connect the client where the research question already comes up, because the agent works best with the surrounding context in view.

• Brand, insights and marketing teams usually start in Claude or ChatGPT. The brief, the saved plan, the approval and the findings all stay in one conversation, and the agent can draft the deck from the report afterward.
• Product teams and developers start in Claude Code, Codex or Cursor. A product manager writing a spec can ask whether shoppers understand a new subscription option and get cited findings beside the code. In Claude Code, the plugin’s skills give the agent a step-by-step workflow for briefing, fielding and analysis.
• Teams building their own agents start with the stdio package or the REST API, and register webhooks so the application learns when interviews finish.

Connecting a second client later changes nothing on the research side. The studies, the reports and the research library belong to the account, so a study briefed in Claude can be analyzed from Codex the next morning.

How do you verify the connection?

Ask the agent: “List my User Intuition studies.” An answer scoped to your account confirms two things at once: sign-in worked, and the research tools are loaded in the client. On a new account the list is empty, and that empty list is still a successful check.

A second useful check is a read that touches the catalog. Ask which languages and panel countries a study can use. The agent reads them from the server, so the answer reflects what your account can field today.

What if the research tools don’t appear?

Clients load MCP tools when a session starts, and hosted connections need a finished sign-in. Check each of these, then ask the agent to list your studies again.

• Start a new session. A server added mid-session usually shows up in the next one. In Cursor, restart or open a new chat; in Codex, start a new task.
• Finish the sign-in. In Claude Code, run /mcp and sign in to the server. In Codex, run codex mcp login userintuition. In Claude and ChatGPT, reconnect the connector and complete the sign-in screen.
• Check the key on a stdio setup. The local package reads USERINTUITION_API_KEY from the environment the client starts it in. With the CLI installed, userintuition-mcp auth status shows whether a sign-in or a saved key is in place.
• Check the server URL. Hosted connections use https://mcp.userintuition.ai/mcp typed exactly, or the profile URL you chose.

How do you test the setup before a real study?

Run a small study with your own participants before you brief a real one, and make two or three teammates the participants. Ask the agent for a short voice study on a low-stakes topic, such as how the team chooses where to order lunch, then have it invite your teammates by email. Each teammate takes the interview from the link in the invitation.

That one run exercises the whole loop: the planning questions, the saved plan, the invitations, the interview itself and the cited report. It also shows your team what a participant experiences before real customers do.

What should your first brief look like?

Write the brief the way you would brief a senior researcher: the decision, the audience, the country, and what you need to learn. A first brief that works in any connected client:

“I want to understand why US coffee drinkers switch between pod and whole-bean brands. Plan an in-depth interview study with the User Intuition panel, ask about the last time they switched, and show me the saved plan and a panel estimate before anything launches.”

Two choices belong to you rather than the agent. The first is who takes part: a vetted 4M+ consumer panel, with B2B audiences recruited to order per study, or your own customers invited by email or link. If the brief leaves recruiting open, the agent asks. The second is the study type. In-depth interviews suit open questions like brand switching, concept tests take an image of the pack or ad, and prototype tests take a live link in voice or video.

Other first briefs that teams start with: a reaction study on a new front-of-pack claim, a shopper-trip study on why a basket ends at a competitor, or a message study comparing two launch emails for a skincare line. Synthetic respondents are also available for a directional first read, in chat, before you field real interviews.

The agent shows you the saved plan and a cost estimate before anything is spent; in Claude you approve with one click, elsewhere you confirm in the conversation. The MCP research API reference walks through each stage from there, up to the cited report.

What can you run once your agent is connected?

• A concept or pack test: run a concept test from an AI agent with images of each pack or product idea.
• A tagline or message test: test taglines with real participants before a campaign locks.
• A churn study: run a churn study from your agent with your own former customers.
• Consumer research through code: the consumer research API guide shows the same jobs as API calls.

Where do you go next?

The research API and MCP server for AI agents page covers the platform as a whole, and the MCP tool reference lists every parameter. The MCP quickstart in the docs holds the configuration files for every client.