How to use Research Helper MCP

Check what you need

A Crowdmines account and an AI platform.

A Crowdmines account
In an organization. Research Helper MCP uses your organization’s credits, as the website does, with no connection or per-call fee. New to Crowdmines? Request a trial first.
An AI platform
The steps below cover claude.ai, Claude Code and Codex CLI. For ChatGPT and other AI platforms, use the last tab in the next step.

Set-up and AI platform coverage are still widening; the statistics are the production engine.

Connect your AI platform

You only do this once. Your browser opens Crowdmines, where you sign in, complete MFA, choose one organization and approve the permissions. It shows as Crowdmines in your AI platform.

Add the server

$ claude mcp add --transport http --scope user crowdmines https://crowdmines.ai/mcp

Sign in
Your browser opens Crowdmines: sign in, verify, choose the organization and approve.
```
$ claude mcp login crowdmines
```
Check the connection
Then ask Claude to list your Crowdmines datasets.
```
$ claude mcp get crowdmines
```

To check it worked, ask your AI platform: What Crowdmines datasets do I have? Your connections are listed, and can be ended, in Account, Connected apps.

Bring your data

Upload a survey and questionnaire from the chat, or work with data already on Crowdmines.

Your file stays private
Your survey file goes straight to Crowdmines, not through your AI platform. Your AI platform only reads the tables and findings you ask for. Attaching the file to the chat would send it to your AI platform, so upload it as shown here.
In claude.ai
Claude opens an upload card where you choose an SPSS .sav file and, optionally, its questionnaire. Up to 200 MiB.
In Claude Code or Codex CLI
Upload .sav, .csv, .xlsx or .xls files, up to 200 MiB.
Already on Crowdmines
Ask for your datasets and pick one. A connection works with one organization.
Give it time
Crowdmines reads the survey before you can ask about it. With a questionnaire this can take up to 30 minutes, so wait rather than uploading it again.
Weighting
Weight your data on the Crowdmines platform. Research Helper MCP then runs on weighted or unweighted data.

Check what you have

Ask for a summary before you run anything: its questions, answer options, sample size and weights.

Give me an overview of this study and what I can analyze.

Ask your research questions

Start from your research questions, one or several at once, the way you would brief a colleague, and Research Helper plans the steps. You can still go table by table. Copy a prompt, then change the names to your own.

Ask several at once
How does purchase intent differ by region? Which age groups are most satisfied? What should the marketing team focus on?
Answer a research question
How does purchase intent differ by age and region?
Compare groups
Compare satisfaction by age group and region, with significance.
Prepare variables
Show top-2-box for brand awareness and NPS by segment.
Group answers
Group Q7 into low (1–4), mid (5–7) and high (8–10), then cross-tab it by gender.
Weighted and unweighted
Run this weighted by the study weight, and show me unweighted too.
Filter
Only respondents aged 18 to 34: cross-tab Q4 by region.
Go one table at a time
Cross Q5 by Q12, column percentages, sig-tested at 95%.
Find what matters
What are the most important findings for the marketing team?
Chart it
Chart purchase intent by region.
Write it up
Write a short report on the top five findings.
Build an output
Turn these findings into a dashboard and a PowerPoint report.

What a session looks like in Claude Code

claude — ~/studies/brand-tracker

> Cross-tab brand awareness by region, with significance. ● crowdmines - find_variables (MCP)(question: "brand awareness")  └ Q4  Have you heard of Northwind?  single choice● crowdmines - run_tabs (MCP)(rows: ["Q4"], banner: "REGION", alpha: 0.05)  └ succeeded · 1 table · base 1,000                Total   North A   South B   West C  Aware          58%     66% C     61% C     47%  Not aware      42%     34%       39%       53% AB  Base         1,000     340       330       330 ● crowdmines - get_findings (MCP)  └ difference     North > West   +19 pts   h 0.39  └ difference     South > West   +14 pts   h 0.28  └ no_difference  North, South ● North and South are both ahead of West on awareness, by 19 and 14 points at the 95% level. North and South do not differ.

Claude Code sessionExample data

Review and download

The numbers come from Crowdmines, not the model, and every table can be downloaded.

In claude.ai
Tables and charts appear in the conversation, with a button that opens the table’s Excel workbook.
In Claude Code or Codex CLI
Tables show as text, with a link to the workbook. Charts come as image links that last an hour.
Same engine as the website
The tables, letters, bases and footnotes Research Helper shows on the platform.
Every comparison classified
Findings list each pair as a difference with its gap and effect size (Cohen’s h or d), no difference, or untested.
One table per question
A multi-select question comes back as one table, one row per option, the way a tab book prints it.
Results outlive the connection
Reading a table never reruns it, and retrying a request does not create a second job.

Getting better answers

What to say so you get the table you meant.

Ask everything at once: Put all your research questions in one message. Research Helper plans the analysis for each.
State the question and the audience: “What drives intent among 18-34s?” works better than “analyze this.”
Name the base: Say who the question is about, such as current customers only or all respondents.
Say how you want to compare: Tell it which groups to compare, such as age, region or brand use.
Say weighted or unweighted: Ask for weighted, unweighted or both when it matters for your reporting.
Build up in steps: Start with an overview, then narrow to the segments that look interesting.
Check the numbers you will publish: Your AI platform writes the words around the tables, so open the table behind any number before it goes into a client deliverable.
Ask why: If a variable choice or recode looks surprising, ask it to explain what it did and why.

When something goes wrong

The usual problems, and what fixes them.

Tools listed twice: The plugin and the manual server are both installed, or the connector was added twice. Remove one.
Codex connects but shows no tools: Turn on the protocol feature (codex features enable mcp_2026_07_28), then restart Codex.
claude.ai sign-in fails at once: The connector was set to Register automatically. Remove it and add it again with Use Claude’s published identity.
Claude doesn’t use the connector: It is switched off for that chat. Start a new chat, then choose +, Connectors, and turn Crowdmines on.
Upload in claude.ai fails with “could not resolve host”: Claude tried to upload from its sandbox. Ask it to open the Crowdmines upload card, and choose the file there.
“Survey is still being prepared”: The survey is still being read; with a questionnaire this can take up to 30 minutes. Wait, and don’t upload it again.
Wrong organization: A connection is fixed to one organization. Disconnect, connect again and choose the other.
Connection gone after 30 days: Connections expire after 30 days. Connect again.
Numbers differ from the website: Compare the footnote and filters. Both run the same engine, so the weighting or filters will differ.

Technical details

For setting up other clients, or answering your security team.

Server URL: https://crowdmines.ai/mcp
Transport: Streamable HTTP
Protocol: MCP 2026-07-28
Authorization: OAuth authorization code with PKCE. Clients identify with a client ID metadata document; dynamic client registration is not supported.
Scopes: datasets:read datasets:write analysis:read analysis:run offline_access
Tokens: Access tokens last 5 minutes and refresh by rotation. A connection lasts 30 days, or until you disconnect it.
Organization: One per connection, chosen when you approve it.
Uploads: .sav, .csv, .xlsx or .xls, up to 200 MiB. The claude.ai upload card takes .sav.
Methods: Cross-tab, cross-tab with significance, banner, frequency, describe, top-2-box, NPS, battery means, funnel, penalty-reward.
Significance: Within a banner group, at 95% and 90%, minimum base 30. All adjustable per request.
Filters: Equals, not equals, in, not in, greater, at least, less, at most, between; combined with AND.
Billing: Your organization’s credits, as on the website. No connection or per-call fee; a retried request does not create a second job.

Sign-in is your Crowdmines account, including MFA.
File contents stream directly to Crowdmines, not through the model.
Every read is authorized per request.
The MCP service holds no database, storage or model-provider credentials.

Tools

You ask in plain words and your AI platform picks the tools. Each call is checked against the permissions you approved.

Upload

request_uploadShows the upload card, in clients that draw views.datasets:write
upload_datasetAuthorizes a direct file transfer from your machine.datasets:write
complete_uploadFinalizes a transferred file.datasets:write
abort_uploadCancels an unfinished upload.datasets:write

Explore

list_datasetsDatasets in the connected organization.datasets:read
describe_datasetVariables, batteries, weights and banner options.datasets:read
find_variablesMatches a research question to questions and answer options.datasets:read
related_variablesBattery and multi-select siblings, and recodes.datasets:read
list_bannerVariables that can form banner columns, with their bases.datasets:read

Prepare

derive_variableCreates a new dataset with a derived variable.datasets:write
rebuild_graphRebuilds the survey graph, keeping the data.datasets:write

Tabulate

run_tabsRuns cross-tabs and banners, with significance and filters.analysis:run
get_operationReads a job’s status and results. Polling never charges.per job
get_tableReads a finished table exactly as rendered.analysis:read
get_findingsEvery comparison as data: difference, no difference or untested.analysis:read