Usercall MCP - AI agents that run real user interviews

AI can build products. But it still doesn't talk to users.

Give your AI agents the ability to ask real users why.

Usercall MCP lets AI agents run user interviews via voice or text and return structured insights with themes and verbatim quotes.

Why this exists

AI agents can now build and ship products extremely quickly.

But most agents still rely on synthetic feedback or assumptions about users.

Usercall MCP lets agents gather real qualitative feedback directly from users.

Choose a connection

Recommended: hosted MCP (Claude, ChatGPT, Cursor, Grok Bot)

Add https://mcp.usercall.co as a remote MCP connector / custom connector.

• OAuth sign-in (no API key, no npx)
• Same tools as this package (studies + Research Triggers)
• Docs: app.usercall.co/docs/mcp
• Cursor Directory / Grok Bot: this repo ships .mcp.json so cursor.directory can install the hosted connector. Grok Bot cannot run the local npx package.

This package: local / API-key / machine-to-machine

Use @usercall/mcp over stdio when you want a Bearer API key (scripts, local clients, M2M).

1. Sign in at app.usercall.co → Home → Developer → Create API key
2. Run npx -y @usercall/mcp with USERCALL_API_KEY

Example workflow

Agent: "Why are users confused about onboarding?"

→ create_study
→ share interview_link with users
→ get_study_results

The returned interview_link can be shared with participants through email, Slack, Discord, or in-product prompts.

Example result:

{
  "themes": [
    {
      "name": "Onboarding confusion",
      "summary": "Users struggled to understand the second step.",
      "quotes": [
        "I wasn't sure what the app was asking me to do.",
        "I didn't know I had to verify my email before continuing."
      ]
    },
    {
      "name": "Pricing confusion",
      "summary": "Free plan limits were not clearly communicated.",
      "quotes": ["I wasn't sure if the free plan included analytics."]
    }
  ]
}

How it works

AI Agent

↓

Usercall MCP (hosted OAuth or this stdio package)

↓

Usercall Agent API

↓

Real user interviews

↓

Themes and verbatim quotes returned to the agent

With Research Triggers, the agent can also target users in your product:

Analytics MCP (PostHog, Mixpanel, …) finds a behavior

↓

Usercall MCP creates a study and a paused Research Trigger

↓

You activate it in Usercall

↓

The Usercall SDK invites matching users to an interview right after the behavior

Research Triggers

Analytics tells an agent what users do. Research Triggers let it ask them why.

User:  "Look at our PostHog data and find something worth investigating."

Agent (PostHog MCP):  users who test a study rarely launch one.

Agent (Usercall MCP):
  list_trigger_events()                  → study_tested, study_launched, …
  get_trigger_event_schema("study_tested")
                                         → properties: source, interview_type
                                           traits: plan ("free", "pro"), account_type
  create_study(...)  or  list_studies()
  create_research_trigger({
    study_id, event_name: "study_tested",
    traits: { plan: "free" }, sampling_percent: 25, max_invites_per_day: 10
  })                                     → status: "paused", summary, activation_url

Agent: "I've prepared a Research Trigger. When: study_tested · Audience: plan = free ·
        25% sampled · max 10 invites/day. It's paused. A human activates it here: <activation_url>"

• The Usercall SDK has to be installed. If list_trigger_events returns nothing, call get_trigger_sdk_setup (with your analytics provider and event names) to get the snippet. Coding agents can install it for you.
• Only events Usercall has actually received can be used. Filters are exact matches on event properties or user traits. get_trigger_event_schema shows which field is which.
• Unsupported conditions are rejected, not silently dropped. These include event counts, sequences, absence ("did not do X"), time windows, and not-equals. get_trigger_capabilities returns the full list.

Local install (API key)

1. Get an API key

Sign in at app.usercall.co → Home → Developer → Create API key

2. Add to your MCP client

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "usercall": {
      "command": "npx",
      "args": ["-y", "@usercall/mcp"],
      "env": {
        "USERCALL_API_KEY": "your_key_here"
      }
    }
  }
}

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "usercall": {
      "command": "npx",
      "args": ["-y", "@usercall/mcp"],
      "env": {
        "USERCALL_API_KEY": "your_key_here"
      }
    }
  }
}

For Claude, ChatGPT, or Cursor remote connectors, prefer https://mcp.usercall.co instead of this JSON config.

Restart your MCP client.

3. Ask your agent

Run user interviews to understand why users drop off during onboarding.

Context:
- B2B SaaS product
- 3-step signup flow

Goal:
Identify confusion points and friction.

Target interviews: 5
Language: ko
Interview mode: voice

Show participants this prototype during the interview:
https://www.figma.com/proto/abcd1234/onboarding-flow

The agent will:

1. create a study
2. return an interview link
3. collect responses
4. return the summary (themes, insights, and risks)

Structured tool example

Equivalent create_study tool call:

create_study
key_research_goal: "Understand why users drop off during onboarding"
business_context: "B2B SaaS signup flow"
target_interviews: 5
languages: ["en"]
interview_mode: "voice"

study_media:
  type: "prototype"
  url: "https://www.figma.com/proto/abcd1234/onboarding-flow"
  description: "New onboarding flow concept"

Tools

create_study

Create an interview study when you already know what happened and still need to learn why. Returns study_id and interview_link. Do not share the link yet. Call list_studies first and reuse a study that already asks this question. key_research_goal is required. business_context is optional. One active agent study per account. On 402, surface checkout_url to a human. This does not run the interview. Call simulate_interview next.

One locale turns the language picker off; two or more turn it on. Research goal cannot be changed after create.

study_media (optional). Visual stimulus shown during all interview questions:

• image: Direct image URL (.png, .jpg, .gif, .webp)
• prototype: Figma prototype URL (converted to interactive embed)
• Media is only visible to web participants; phone callers won't see it

update_study

Edit an existing study's slots, interview mode, languages, voice, link context, guide text, questions, or media. Use this after review_study or a failed simulation names a guide change, or when the link is disabled and you are about to share. You cannot change key_research_goal. One locale turns the language picker off. Two or more turn it on. Query params on interview_link are ignored until enable_link_context is true. Call simulate_interview again before sharing.

Pass study_media: null to clear media. The study_media object follows the same schema as in create_study.

get_study_status

Check whether a study is running, analyzing, or complete. running and analyzing mean wait and call this again. Do not treat those payloads as findings. When status is complete, call get_study_results. This does not return themes.

Status values: running · analyzing · complete

Response includes interview progress fields, including completed_interviews and target_interviews.

get_study_results

Read findings after get_study_status is complete. Prefer format=summary for themes, insights, and risks. Use format=full only when a verbatim transcript is required. Empty themes while the study is still running are not a finding.

Summary/full responses include study progress fields and analysis output.

simulate_interview

Dry-run the interview after you create or edit a study, and before any real invite. Omit simulation_id to start. Pass that id to read the result. The start returns immediately with running. Cap is 5 simulations per account per UTC day. A simulation is not a completed interview and does not change completed_interviews. On fail, call update_study, then simulate again.

Omit simulation_id to POST /api/v1/agent/studies/{studyId}/simulations. Pass simulation_id to GET that simulation. The tool does not poll.

review_study

Check the interview guide before sharing it. It reads the guide only. It does not read transcripts and it does not apply edits. It costs 1 credit. On 402, surface checkout_url to a human. Write suggested changes with update_study. Stop after one review unless the guide changed. This is not simulate_interview and it is not get_study_results.

Sends study_id only. It does not send call_ids.

delete_study

Permanently delete a study when it asks the wrong question or you must free the one active agent study. This cannot be undone. To stop new interviews without deleting evidence, call update_study with is_link_disabled true. This does not delete a research trigger.

Research Trigger tools

create_research_trigger

For page-visit triggers, use source: "page_visit" and event_name: "$pageview", with url and optionally dwell_seconds.

Delivery.

• intercept (default) shows the Usercall widget in your product, and the user takes a voice or text interview in the page. The modes come from the study; list_studies returns each study's interview_mode.
• webhook POSTs each matched user to webhook_url, with their user ID, email if known, traits, event properties and a personal interview link. If webhook_secret is set, requests carry an x-usercall-signature HMAC header.
• Only public https URLs are accepted, and the activation page shows the destination before a person activates the trigger.

Safety

• Agents cannot activate triggers. Triggers are always created paused. Calling update_research_trigger with status: "active" returns HTTP 409 and the activation_url. A person has to open that link, review who will be invited, what they will see and the credit cost, and click Activate.
• Changes to an active trigger need re-approval. Changing an active trigger's configuration pauses it again.
• Secret keys are never returned. The ingestion secret key never comes back from any tool.

Example workflow

1. create_study
   key_research_goal: "Why do users drop off during onboarding?"
   business_context: "B2B SaaS, 3-step signup flow"
   target_interviews: 5
   languages: ["ko"]
   interview_mode: "voice"

   → returns { study_id, interview_link }
   (`business_context` is optional; `key_research_goal` alone still creates a study)

2. simulate_interview
   study_id
   → running, simulation_id
   call again with simulation_id
   → pass, fail, or error

3. review_study
   study_id
   → guide check only; write changes with update_study

4. Share interview_link with participants
   (email, Slack, in-product prompt, etc.)

5. get_study_status
   → "analyzing"

6. get_study_results
   → summary: themes, insights, and risks
   use format=full only for a quote

With visual stimulus

1. create_study
   key_research_goal: "Get feedback on new dashboard design"
   business_context: "Redesigning analytics dashboard for power users"
   study_media:
     type: "image"
     url: "https://example.com/dashboard-mockup.png"
     description: "New dashboard design concept"

   → returns { study_id, interview_link }

2. After simulate_interview passes, a human shares interview_link. Participants see the mockup during the interview.

For Figma prototypes, use type: "prototype" with a Figma proto URL.

Requirements

• Node.js 18+
• A valid Usercall API key (local / API-key path only)

Self-hosting / development

pnpm install
pnpm build
USERCALL_API_KEY="your_key_here" pnpm start

Tests and smoke tests:

pnpm test                                   # unit + MCP contract tests
USERCALL_API_KEY="your_key_here" pnpm smoke # creates a real study
USERCALL_API_KEY="your_key_here" SMOKE_STUDY_ID="<uuid>" SMOKE_EVENT_NAME="<observed event>" pnpm smoke:triggers

Official MCP Registry

Usercall is listed on the Official MCP Registry as co.usercall/mcp.

Troubleshooting

Remote Claude / ChatGPT / Cursor connectors should use https://mcp.usercall.co (OAuth). This package is the API-key stdio path.

License

MIT