Claude + MCP for Customer Research: Setup, Workflows, and Common Patterns

Claude Desktop and Claude Code are the highest-volume MCP clients available — which makes the User Intuition + Claude integration the fastest path to conversational, agent-driven customer research without writing custom code. Once connected, Claude can recruit participants, run AI-moderated studies, retrieve findings, and query the Intelligence Hub in response to natural-language prompts. The research workflow stays inside the conversation.

This guide covers setup for both Claude Desktop and Claude Code, five end-to-end workflow examples, common gotchas, and how to get the most from the integration. For the full API surface, see /platform/agentic-research/ and the tool reference at docs.userintuition.ai/mcp-server/overview.

Why Claude + MCP changes customer research

The traditional research workflow requires context-switching: you have a question in your AI assistant, you switch to a research tool to configure a study, you wait, you come back, you paste results into your original conversation. The MCP integration removes every switch.

When the User Intuition MCP server is connected to Claude, the research tools appear in Claude’s native context. Claude can call ask_humans to check a message with a real panel, create_assistant to set up a full moderated study, analyze_transcripts to retrieve findings, and query_intelligence to synthesize answers from past research — all within the same conversation, responding to the same natural-language prompts that drive the rest of the work.

For researchers, this means faster iteration: a positioning question that previously required a 3-day research cycle can get directional signal in an hour. For product teams and AI agents, it means research becomes a callable step in the workflow rather than a separate project. For teams building agent workflows, it means Claude can decide when to commission research, not just execute it when instructed.

Setup: Claude Desktop

Claude Desktop connects to MCP servers via a JSON config file. Locate it at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS (or the equivalent on Windows). Open the file in a text editor and add the userintuition entry to the mcpServers object:

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

Replace ui_sk_your_key_here with your actual API key from the User Intuition dashboard.

Save the file and restart Claude Desktop completely (quit and relaunch — a window refresh is not sufficient). On next launch, the research tools will be available in Claude’s tool context. You can verify the connection by asking Claude: “What research tools do you have available?” — Claude will enumerate the tools it can call.

If you have other MCP servers already configured, add the userintuition entry alongside them in the same mcpServers object. Do not replace the entire config.

Setup: Claude Code

Claude Code (the CLI) manages MCP servers via the claude mcp command. Run:

claude mcp add userintuition -- npx -y @userintuition-ai/mcp

This registers the server. Then set your API key. The simplest approach is exporting it in your shell profile so it is available in all Claude Code sessions:

export USERINTUITION_API_KEY="ui_sk_your_key_here"

Or, if you want it scoped to a specific project:

# In your project root, create .env.local or set in your shell before launching:
export USERINTUITION_API_KEY="ui_sk_your_key_here"
claude  # starts Claude Code with the key available

Verify the connection by starting a Claude Code session and asking: “List the User Intuition tools available.” Claude Code will call the server’s tool-list endpoint and return the 72 tools organized by group.

5 workflow examples that work end-to-end

These prompts are ready to use in Claude Desktop or Claude Code once the MCP server is connected. Each uses real tool names from the User Intuition MCP server.

Workflow 1: Preference check (Human Signal)

Prompt:

I have two headline options for a landing page targeting B2B SaaS buyers.
Option A: "Know what your customers think before you ship"
Option B: "Customer evidence on demand"

Run a preference check with 20 B2B SaaS product managers. Use ask_humans
and call get_results when complete.

Claude will call ask_humans with the question and segment parameters, then poll get_results until fielding completes. Typical turnaround: 1-3 hours.

Workflow 2: Claim test (Human Signal)

Prompt:

Test this claim with 15 enterprise buyers in financial services:
"AI-moderated interviews return findings in 24-48 hours, not 3 weeks."

Rate credibility on a 1-5 scale and capture their reasoning. Use ask_humans.

Workflow 3: Message test (Full moderated study)

Prompt:

Set up a 10-person audio study to test our pricing messaging.
Target: B2B SaaS product leaders, director level and above, company size 50-500.
Language: English.

Discussion guide focus: pricing reaction, value perception, objections.

Call create_assistant with the study configuration, then
bulk_create_invites_from_segment for the audience.
Let me know when it's set up so I can review before activating.

Claude will call create_assistant with the study parameters, then bulk_create_invites_from_segment with the targeting filters. It will confirm the setup is complete and wait for your approval before fielding starts.

Workflow 4: Full moderated study, end-to-end

Prompt:

Run a full churn analysis study autonomously:
- Target: former customers who churned in the last 90 days (I'll provide a segment ID)
- Mode: audio interviews
- Sample: 12 participants
- Discussion guide: churn reasons, what they switched to, what would have kept them

Create the assistant, invite the segment, wait for completion, analyze transcripts,
and generate a report. Update me at each major step.

Claude will orchestrate the full sequence: create_assistant → bulk_create_invites_from_segment → monitor via list_calls → analyze_transcripts → generate_report. The workflow runs over 2-4 hours with Claude checking in at each phase transition.

Workflow 5: Intelligence Hub query

Prompt:

Query the Intelligence Hub for: "What objections come up most often
when enterprise buyers evaluate our pricing?" Return findings with
source references so I can see which studies they came from.

Use query_intelligence.

Claude calls query_intelligence with the natural-language query and returns synthesized findings across all completed studies in the workspace. This pattern requires existing research in the workspace; the more studies accumulated, the higher-confidence the synthesis.

Common gotchas

These are the issues that trip up most new integrations. They are all manageable once documented.

Always read before update

update_assistant does not work correctly without a prior get_assistant read. Before calling update_assistant, call get_assistant to retrieve the current state. Pass the retrieved values plus your modifications to update_assistant. Calling update_assistant without first reading can produce unexpected results on fields you did not intend to change.

Correct pattern:

get_assistant(assistant_id="...")
→ [Claude reads current state]
→ update_assistant(assistant_id="...", {current_fields + your_changes})

screener_questions REPLACES, not merges

When update_assistant receives a screener_questions parameter, the value completely replaces the existing screener list. If the study has three screener questions and you add a fourth via update_assistant, you must pass all four in the screener_questions array — otherwise the study will have only the new question.

Correct pattern:

get_assistant → retrieve current screener_questions list
→ append new question to the list
→ update_assistant with the complete combined list

No auto-screeners on study create

create_assistant does not automatically generate screener questions based on the audience parameters you provide. If you want screeners, add screener_questions explicitly in the create_assistant call (or in a subsequent update_assistant call after reviewing the initial configuration). Claude will not invent screeners for you unless you ask it to propose them and confirm before creating.

Authentication uses an environment variable, not a request header

The API key is set as the USERINTUITION_API_KEY environment variable in the MCP server config. It is not passed as a request header. If you are seeing authentication errors, verify the environment variable is set in the env block of your Claude Desktop config or exported in your shell before starting Claude Code.

No webhooks

The User Intuition MCP server does not use webhooks for completion notification. To check whether interviews are complete, poll list_calls or get_assistant_report until the status indicates completion. In a Claude conversation, asking Claude to “check status and let me know when complete” works — Claude will poll at a sensible interval and notify you when the study is done.

How does User Intuition handle Claude-driven research?

User Intuition built its MCP server specifically for the Claude + MCP integration pattern. The 72-tool surface was designed to support conversational orchestration: Claude can discover available tools from the server manifest, call them in natural-language-driven sequences, and chain results across a multi-step workflow without custom integration code.

The Claude integration unlocks three specific capabilities that matter for research. First, the AI moderator and Claude’s language capabilities operate in the same conversation: Claude configures a study based on your description, the AI moderator runs the interviews, and Claude synthesizes findings — the handoffs are seamless. Second, query_intelligence in the Intelligence Hub is optimized for natural-language queries exactly like the ones you’d ask in a Claude conversation — the tool was designed for this pattern. Third, the generate_powerpoint and generate_report tools produce formatted outputs that Claude can summarize, contextualize, and prepare for distribution within the same session.

Authentication is a single ui_sk_ key with no scopes, no webhook setup, and no additional configuration. The Starter plan includes three free interviews with no card required — the full 72-tool surface is available from day one. The Professional plan ($999/month) includes 50 credits monthly at $20/credit for audio interviews. Participant panel coverage spans 4M+ participants across B2B and B2C segments in 50+ languages. Studies at the standard 20-50 participant range complete in 24-48 hours from recruit to analyzed findings. Full platform documentation at /platform/agentic-research/.

Get a key and try it

1. Sign up at userintuition.ai — three free interviews included, no card required
2. Generate a ui_sk_ key from your dashboard
3. Add the config block above to Claude Desktop or run the claude mcp add command
4. Restart Claude Desktop (or start a new Claude Code session)
5. Ask Claude: “Run a quick preference check with 10 B2B SaaS buyers: which headline do they prefer, A or B?”

Full documentation including tool schemas, parameter references, and integration examples: docs.userintuition.ai/mcp-server/overview.