The User Intuition MCP research API turns a qualitative study into a sequence of tools an AI agent can call: create, plan, estimate, launch, monitor, report and deliver. This guide is the reference for that sequence. It maps the tools by job, follows one study from brief to evidence, shows exactly where your approval sits, and describes what each stage returns. For exact inputs and schemas, use the MCP tool reference.
What is the User Intuition MCP server?
The User Intuition MCP server is the connection that exposes the MCP research API to an agent. An MCP server makes outside capabilities discoverable: when Claude, ChatGPT, Claude Code, Codex or Cursor connects, the agent reads the available tools and calls them from context. User Intuition uses that interface for qualitative research with real participants. The agent can create a study from a brief, relay the planner’s questions, show you the saved plan and estimate, launch on your approval, and return findings cited to the quotes behind them. User Intuition’s AI moderator runs 20–30 minute interviews that ladder 5–7 levels deep, by voice, chat or video in 80+ languages. The same server produces reports, decks and highlight videos from finished studies. It is the research API and MCP server for AI agents, reachable over hosted OAuth or a local process with an API key.
What does the MCP research API cover?
Capabilities map to the jobs of a research project. The tool reference lists every operation.
| Job | What the agent does |
|---|---|
| Plan | Create a study, send your brief through the planning step, relay the planner’s questions, and show you the saved plan |
| Recruit | Look up countries, languages and targeting attributes; estimate cost and timeline; launch with the vetted 4M+ consumer panel, request feasibility for hard-to-reach audiences, or invite your own participants by email or link |
| Field | Run in-depth interviews, concept tests with an image, and prototype tests with a live link; pause or resume fielding |
| Monitor | Report completed interviews and quality counts; open transcripts, recordings, quality ratings and screener answers; update participants and send rewards |
| Analyze | Generate and read reports with prevalence (n of N) and verbatim quotes linked to the source interview |
| Search | Answer questions from earlier studies with citations to the source interviews |
| Deliver | Generate reports, PowerPoint decks and highlight videos; edit and download them |
| Automate | Receive webhook events for completed interviews, finished studies and ready reports |
How does a study move from question to evidence?
Every study follows the same eight steps, whichever client the agent runs in. Each step names the tool that does the work.
- Choose who takes part.
create_studyrecords the study and your recruiting choice: a vetted 4M+ consumer panel, with B2B audiences recruited to order per study, or your own participants. The agent asks if the brief leaves it open. - Plan.
customize_studysends your brief through the planning step. The agent relays the planner’s questions and sends your answers back. - Review the saved plan.
get_studyreturns the learning goals, discussion guide, screener questions and audience that participants will meet. - Estimate. For a panel study,
estimate_panelquotes the cost, your balance and the timeline for the country and audience you chose. - Approve and launch. You approve the exact plan and estimate.
launch_panelfields a panel study against that estimate, andcreate_participantsinvites your own participants. - Monitor.
get_fielding_progressreports completed interviews and quality counts, and your team can watch any interview in Live view. - Read the report.
get_study_reportreturns findings grouped by learning goal, with prevalence counts, quotes, evidence coverage and recommended next steps. - Deliver or follow up. The agent turns the report into a deck or highlight video, or plans the next study from the open questions.
The sections below show what each step calls and returns.
How do agents reach the server?
Connect it in Claude, ChatGPT, Codex or Cursor in a few minutes (setup). Every client reaches the same tools at https://mcp.userintuition.ai/mcp, over one of two transports:
- Hosted Streamable HTTP signs in with OAuth for interactive clients and accepts an organization API key for headless ones.
- Local stdio runs the npm package as a local process with your API key in the environment.
Core, read-only and compact profiles trim the catalog for agents that need fewer tools. Every request is scoped to your account, so an agent sees your studies and only yours.
How does the agent design the study?
You choose how to recruit when the study is created. The agent creates the study with that choice, then sends your brief through the planning step. Study types are in-depth interviews by voice, chat or video, concept tests with an image, and prototype tests with a live link by voice or video.
create_study({"name":"Oat milk pack claims","recruiting_method":"panel","study_type":"concept-test"})
customize_study({"study_id":"<study-id>","message":"Test three front-of-pack claims for our oat milk with US shoppers who bought plant-based milk in the last month. Explore which claim they believe, what it leads them to expect, and what would make them switch brands. <attach the pack image>"})
get_study({"study_id":"<study-id>"})
The planning step answers with questions about anything the brief leaves open: who qualifies, which screener questions to ask, which learning goals matter most, and how to present the concept. The agent relays those questions to you and sends your answers back. If you hand design choices to the agent, the planner proceeds on stated assumptions and the agent reports them for your review.
get_study returns the saved plan: learning goals, the discussion guide, screener questions and the audience. You read that plan before anything fields, and you can ask the agent to change it with another planning turn. Synthetic respondents are also available for a directional first read, in chat, before you commit to real interviews.
How do approval and launch work?
Approval is where the agent hands control back to you, because a launch commits budget. Nothing is spent until you approve the saved plan and its estimate.
For a panel study, the agent calls estimate_panel with the country, target and incidence you choose. The estimate returns the cost per interview and in total, your available balance, and the expected timeline for that audience and country. Most panel studies fill within 24 hours.
estimate_panel({"study_id":"<study-id>","target":25,"incident_rate":50,"country_code":"US"})
launch_panel({"study_id":"<study-id>","target":25,"incident_rate":50,"country_code":"US","estimate_id":"<estimate-id>"})
How you approve depends on the client:
- Claude. Claude supports MCP Apps, so
launch_panelopens a review card with the plan and the quote side by side, and you approve with one click. - ChatGPT, Claude Code, Codex and Cursor. The agent shows the plan and estimate, and you confirm in the conversation.
- The REST API. Your code launches against a saved estimate.
In every case the launch names the estimate you reviewed, and the first launch_panel call only previews. The price you saw is the price of the study that fields, and if the plan or the price changes, the agent brings you a fresh estimate to review. A retried launch with the same idempotency key cannot field the study twice.
For your own participants, the agent sends invitations by email or link only after you approve the saved plan, and each invitee receives an interview link. After launch, the agent can pause or resume fielding on request.
What happens while interviews run?
get_fielding_progress reports completed interviews and quality counts against the target. Because the study ID carries everything, the agent can check in from any later session, so a long study never ties up the conversation.
Your team can also watch fielding as it happens. Live view lets anyone on the team join an interview in progress, in voice, video or chat, and send the moderator a silent instruction to probe a detail. Every instruction is logged in the transcript.
Software that should react without being asked can subscribe to webhook events such as interview.completed, study.completed and report.ready through create_webhook. Whether to poll or subscribe is a design choice covered in MCP vs REST for user research.
What does get_study_report return?
get_study_report({"study_id":"<study-id>","view":"study_findings"})
The default overview returns the summary, interview count, freshness and the sections available. Each section can then be fetched on its own:
- Study findings: findings grouped by learning goal, each with a prevalence count (n of N) and verbatim quotes linked to the source interview and, for voice and video studies, to the moment in the recording.
- Participant profiles: who took part, summarized from profile and screener answers.
- Participant responses: each interview’s takeaways against every learning goal.
- Evidence coverage: how well the evidence supports each learning goal, and the most important open questions.
- Recommended next steps: up to two follow-ups, each tied to an open question in evidence coverage.
Fetching sections one at a time keeps the agent’s context small: it can read the overview, pull only the findings for the learning goal you asked about, and leave the rest until needed.
How should an agent read a finding?
A finding is built to be checked, so an agent should read it in three passes.
First, the prevalence count. A finding held by most of the interviews in a study carries different weight from one raised by a handful, and the count says which it is. Second, the quotes. Each quote links to the interview it came from and, for recorded studies, to the moment in the recording, so a person can hear the participant say it before the finding drives a decision. Third, evidence coverage. A learning goal with thin coverage is a signal to ask a follow-up or run a narrower study, and the recommended next steps name the open question that would close it.
That order keeps an agent from presenting a one-off remark as a pattern. To see a finished report with real findings and recordings, explore an example study.
What deliverables can the agent produce?
Before briefing anything new, an agent can also ask questions across past research and plan a study only for what earlier work leaves open.
Finished studies turn into deliverables in the same conversation. generate_research_output creates a written report, a native PowerPoint deck, or a highlights video assembled from the report’s verified quotes, with original participant audio where it exists. edit_research_output revises a report or deck in place, for example “shorten the summary and keep every citation,” and download_research_output returns Markdown, PPTX or PDF files. Over the REST API, interview exports deliver transcripts as CSV or JSONL files.
What do workflow prompts, skills and the CLI add?
The server publishes workflow prompts that a client discovers on connect, alongside the tools. In Claude Code, the user-intuition-research plugin includes the hosted server and adds ten skills, described in the skills library. For scripts and CI jobs, the userintuition-mcp CLI estimates panels, inspects fielding and reads results with JSON on stdout, while a paid panel launch runs from an MCP client or the REST launch-panel endpoint. The CLI overview has the commands.
Which workflow fits your question?
A focused decision. A narrow brief for a pack, claim or concept, with the audience and the decision stated. See how to run a concept test from an AI agent.
A message decision. Two or three lines tested before a campaign locks. See tagline testing from an AI agent.
A study with your own customers. Former or current customers invited by email, for questions such as why people leave. See running a churn study from an AI agent.
What makes the evidence trustworthy?
User Intuition’s AI moderator runs 20–30 minute interviews that ladder 5–7 levels deep, from what a participant did to why it mattered to them, so findings carry reasons rather than one-word answers. Five controls from User Intuition’s quality controls protect every interview:
- A member of the User Intuition team listens to every call with a panelist.
- Panelists pass a video identity interview before they join.
- Voice interviews are screened with AI voice detection for AI-generated answers.
- Your team can watch any interview live with Live view.
- Only interviews that pass length, depth and coverage scoring are billed, and you can reject any interview.
User Intuition interviews hold 98% participant satisfaction, and the platform is rated 4.9/5 on G2 and 5/5 on Capterra.
How do MCP and REST differ?
Both reach the same studies and reports: hosted MCP lets an agent discover tools from context, while REST suits code that runs known steps. MCP vs REST for user research covers when to use which, and the developer integration guide walks the same study through REST with code.
Voice interviews cost $30 each with no separate fee for the MCP connection, and signup includes 3 free interviews with your own participants; panel recruiting is billed separately (pricing).