MCP (AI)

What the QUESTIONSTAR MCP is — in one minute

The Model Context Protocol (MCP) is an open standard that connects AI assistants to live data and tools. Think of it as USB-C for AI: any MCP-compatible AI client (Claude Desktop, ChatGPT, Cursor, Zed and more) can plug into any MCP-compatible server and use what it offers — read access, write access, structured queries — without anyone having to program bespoke integrations.

The QUESTIONSTAR MCP server makes your surveys, responses, respondents and exports available to these AI clients. When you ask the AI a question about your data, it doesn't rely on training knowledge or an outdated copy — it calls the QUESTIONSTAR MCP, reads the live data and answers based on what it actually finds.

A real-world example: before and after

Imagine you run a customer satisfaction survey every month. Each month you put together a summary for the management team.

Before MCP — the typical workflow

QUESTIONSTAR's built-in analytics do most of the work: charts and distributions, crosstabs, Word Cloud and TF-IDF on open answers, even ready-made Word/PDF exports with your own notes. So the "before MCP" workflow has nothing to do with calculating — QS handles that. It's about the human work that sits on top of the analytics:

1. Open the survey in QUESTIONSTAR. Look at the report — numbers, charts and the Word Cloud / TF-IDF for open questions are right there.

2. Read 50–200 open answers yourself to get a sense of what respondents are really saying — the Word Cloud hints at themes, but you have to do the interpreting.

3. Compare with the previous period by eye — pull last month's numbers, set them next to the current ones, keep the deltas in your head.

4. Look for cross-comparisons the standard report doesn't provide for — "Is the cluster of negative comments concentrated in a particular demographic group?" or "Do respondents who complain about price also rate support lower?" Usually a manual filter pass.

5. Write the narrative text. QS exports your charts and notes to Word; most people still rewrite the prose anyway — for the audience in question (short executive summary vs. 5-point team digest vs. detailed board memo).

6. Polish, send.

Realistic duration: 30–60 minutes per recurring report. Most of that is interpretation, narrative writing and cross-comparisons the standard report can't anticipate.

With MCP — the same task

You open your AI client and type:

Compare this month's customer satisfaction with last month. Pull the recurring themes from the open answers. Flag demographic patterns in the negative comments. Write a one-page executive summary for the management meeting — 3 strengths, 3 concerns, 2-3 verbatim quotes.

The AI reads the data live via the MCP, calculates the period delta, classifies the sentiment of the comments, finds the cross-comparison between negative comments and respondent demographics (exactly the kind of insight a fixed report draft doesn't produce automatically) and drafts the text.

Realistic duration: 3–5 minutes.

The real point isn't the time saved on this one task — it's that after a one-time setup you can ask at any time for interpretations, summaries in several languages, non-standard period or segment comparisons, correlations between text answers and demographics — anything the standard template doesn't provide for. Each of these used to be its own piece of manual work; today it's a single sentence.

What the MCP can do — an overview

The QUESTIONSTAR MCP currently provides over 20 tools. You don't need to memorize any tool names — the AI picks the right one for your request. All you need to know is what kinds of questions you can even ask. Below is a thematically organized overview with one or two real prompts per topic area and a description of what happens.

Which surveys do you have, and how are they structured?

Tools used: list_surveys, get_survey, get_survey_result_items

Example prompt:

List my surveys and tell me which ones received new responses in the last 7 days.

What the AI does: calls list_surveys to fetch all surveys, then get_survey_last_response per survey to check the last response date, filters to the last 7 days and shows the result as a short table. With 50 surveys that takes a few seconds — clicking through the dashboard manually would cost you 5 minutes.

Another example:

Open my Customer Satisfaction survey and tell me which questions it contains.

What the AI does: calls list_surveys to find the survey by name, then get_survey with the survey code to load the schema. Returns a numbered list of the questions with type (radio, text, matrix, etc.) and answer options.

Handy before any analytical request — it lets the AI ground its later answers in the actual schema, not in a guess. Also works well to remind yourself what's in a survey you haven't looked at in months.

Reading actual responses

Tools used: get_result_cases, get_result_case, get_result_case_complete

Example prompt:

Show me the 5 most recent responses to my Employee Survey. Per response, the rating questions and the open comment.

What the AI does: finds the survey, fetches the most recent cases via get_result_cases, loads the complete answer set per case via get_result_case_complete and shows it all as a compact table. The AI automatically leaves out metadata (timestamps, technical IDs) unless you explicitly ask for it.

Another example:

Find all responses where the satisfaction rating was 1 or 2, and add the open comment. Group by the respondent's department.

What the AI does: loads all cases, filters in memory by the rating value, groups by department and gives you exactly the dissatisfied responses with their comments — your to-do list in a single prompt.

Filtering and exploring responses

Tools used: get_survey_cases_by_dates, get_survey_devices, get_survey_sources, get_survey_qualified, get_survey_disqualified

Example prompts:

How many people opened my NPS survey last month on mobile vs. desktop?

My citizen survey ran for 3 weeks. Show me the response volume per day. Was there a peak right after the reminder email went out?

Of the 800 people who started the survey — how many were disqualified by the screener? How are the disqualification reasons distributed?

Each of these requests would otherwise be a click marathon through the QUESTIONSTAR dashboard plus an export. Via the MCP it's one sentence and a 2-second answer.

Creating respondents and responses programmatically

Tools used: create_respondent, get_respondent, create_result_case, update_result_case

These tools write into QUESTIONSTAR, not just out of it. Real situations where that's useful:

• Transferring a paper survey. You ran a paper survey at an event. Give the AI the photos or the typed-up table — it creates each respondent and their answers as a case via create_result_case.
• Handover from the CRM into the survey. Your CRM holds the contact list for the next wave. Ask the AI to add these contacts as respondents — no manual CSV upload.
• Correcting a response. A participant emails you that they'd like to refine their answer. The AI finds the case via get_respondent and updates it via update_result_case.

Example prompt:

Here are 12 paper responses I just typed up (CSV pasted in the chat). Add them to the survey Employee-Survey-Q4-Paper. Take the names from column A as identifiers.

What the AI does: parses your input, calls create_respondent per name, then create_result_case with the parsed answers per respondent, and reports back how many entries were successfully created.

Data exports

Tools used: get_respondents_pipeline_csv, get_respondents_pipeline_xlsx, get_respondents_pipeline_spss

Example prompts:

Export my Customer Satisfaction survey to SPSS, with value labels for the rating scales.

Give me a CSV with all responses to the Employee Survey from October 2025 — only the rating questions, with German question labels in the header row.

The AI calls the appropriate pipeline tool and delivers a downloadable file. For SPSS exports, the value labels are preserved so the file can be used directly in SPSS — no import wizard needed.

Unexpectedly useful

The tour above covers the mechanics — reading, filtering, exporting. The truly powerful part of the MCP comes after: when you have the AI not just retrieve, but analyze and interpret. What follows is three levels — from quick everyday use to research methods.

Level 1 — Quick summaries and reports — no statistical background needed

This level replaces routine report work that eats up hours today. You don't need any statistical knowledge. The AI does the synthesis, you do the fine-tuning.

Example 1 — Executive summary for management:

Write a one-page executive summary of the Q3 2025 Customer Satisfaction survey for the executive team. Highlight the key trends, 3 strengths and 3 urgent areas for action. Quote 2 verbatim customer voices per area for action.

What you get back: a fully written 1-page text with concrete percentages, three highlighted strengths, three specific areas for action and short, representative quotes. Often 80-90% ready to send — you polish the wording and send the text.

Example 2 — Comparison between periods:

Compare the NPS results between Q2 and Q3. Has the score shifted? In which segments most strongly? For that, pull themes from open answers in Q3 that might explain the change.

What you get back: the score delta with confidence interval, the breakdown per segment and 2-3 themes from the open answers that point to the causes. Often more useful than the plain NPS dashboard — it tells you the why, not just the what.

Example 3 — Team-specific analysis:

My Employee Survey has a filter field "Department". For each department, give me a 3-sentence summary: top 2 strengths, top 2 concerns, one verbatim quote. Formatted as a list.

What you get back: a mini summary per department, ready to paste into the 1:1s with the department heads. A personalized view that used to require a manual Excel filter pass per team.

Level 2 — Interpretation and pattern recognition

QUESTIONSTAR delivers crosstabs, t-tests and correlation coefficients out of the box — you can click most of them together right in the "Reports" area. What standard reports don't do:

• explain to you what the numbers actually mean,
• systematically search all question combinations to surface insights you'd never have gone looking for on purpose,
• include open-ended responses in the correlation search,
• point you toward natural respondent segments without you having to filter manually.

This is exactly where the MCP comes in — as an additional layer on top of the analytics QS already gives you.

Example 1 — Plain-language interpretation of a crosstab QS has already computed:

Look at the crosstab Department × Overall Satisfaction from my employee survey. What do the numbers mean in practical terms? Which department deserves immediate attention, which can serve as a model, and how likely is it that the observed differences are due to chance?

What you get back: a plain-language paragraph on which departments are genuinely relevant, which are more likely to be noise and what a concrete action plan could look like. The χ² test is already in your QS report — the AI translates the numbers into a story and a recommendation. The same applies to t-tests and any crosstab the standard report shows: you supply the table, the AI supplies the interpretation.

Example 2 — Automatic correlation search across all questions, including open-ended answers:

Go through all pairs of questions in the customer satisfaction survey and find the 5 strongest relationships — including unexpected pairs. Include the open-ended answers: does what someone writes in a given question relate to their answers in other questions?

What you get back: a sorted list of question pairs with correlation coefficients and — this is the part the native correlation function in QS can't do — cross-connections that involve open-ended answers. Something like: "Respondents who use the words 'too expensive' in their open comment rate the questions on technical quality on average 1.2 points higher than the sample mean. So the price complaint has nothing to do with quality concerns."

In QS you could correlate each pair of closed questions individually — with 30 questions that's 435 pairs to click through. The AI does it in a single pass, highlighting only the most meaningful findings — and it includes the open-ended answers, which a classic correlation tool can't.

Example 3 — Comparing group means (ANOVA — not included in the standard QS report):

My survey has a "Location" filter field. Compare the average satisfaction score by location. Is the difference statistically meaningful? Which locations lie clearly above and which below the overall mean?

What you get back: means and standard deviations per location, a one-way ANOVA with F-statistic and p-value, plus post-hoc pairwise comparisons (Tukey HSD) so you know which pairs of locations actually differ. ANOVA isn't among the clickable tests in the standard QS report — via the MCP you have it at hand.

Example 4 — Detecting segments:

Look at the response patterns of all respondents. Are there 2-3 natural groups? Describe each group in one sentence: who they are and how they differ from the others.

What you get back: 2-3 respondent profiles in plain language — something like "Group A (35%): consistently high satisfaction across all questions, long-standing customers from larger companies. Group B (28%): mid-range scores with recurring complaints about the onboarding pace. Group C (37%): mixed pattern, no clear profile."

In QS you'd have to manually set up filter combinations, jump back and forth between sub-reports and eyeball where the differences lie. The AI does it systematically and points you straight at the segments it found. Automatic segmentation isn't in the standard QS toolkit — and even if you already know your segments, a one-sentence description per segment is qualitative summary work that a market research consultant used to handle.

Level 3 — Research-grade analysis

This level is for anyone doing genuine research — doctorates, market research consulting — or who wants methods the standard QUESTIONSTAR report doesn't cover.

Example 1 — Principal component analysis (PCA):

Run a PCA over the 12 Likert items in the "Leadership evaluation" block. How many components have an eigenvalue > 1? Show me the loading matrix after varimax rotation. Suggest interpretable names for the latent constructs.

What you get back: the scree-plot data, eigenvalues, the rotated loading matrix, and the AI's interpretation of which items together form, e.g., "clarity of communication", "professional competence", "empathy". Useful for validating scales and for reducing 12 questions to 3 substantively robust indices.

Example 2 — K-means clustering / segmentation:

Cluster my customer respondents based on their response patterns to the 6 satisfaction questions. Try k=2, 3, 4, 5. Which k seems most sensible? For the chosen k, describe each cluster in plain language.

What you get back: the within-cluster sum of squares per k (data for the elbow plot), a recommendation for a k, the cluster centroids, and a paragraph per cluster — something like "Cluster 2 (n=87): generally satisfied with the product, but consistently low scores on support response time. Average customer relationship length 4+ years." Segmentation that used to cost a market research consultant a week.

Example 3 — Topic modeling on open-ended answers:

I have about 300 open-ended answers to the question "What could we do better?". Identify the 5 main themes. Per theme: a label, an estimated frequency, and 2-3 representative quotes.

What you get back: 5 themes with frequencies, such as "price concerns (38%, 114 mentions)", "delivery time (22%, 66 mentions)", plus the most representative quotes per theme. Coding of qualitative data that traditionally costs a research assistant a whole day.

Example 4 — Sentiment analysis:

For each open comment in the customer satisfaction survey, classify the sentiment as positive / neutral / negative. Show me the overall distribution — and broken down by the score the respondent gave in the closed question. Do the comment's sentiment and the numeric score match up?

What you get back: a sentiment label per comment, the overall distribution, a contingency between sentiment and rating, plus an observation like "15% of respondents who gave the highest numeric score nonetheless left a comment classified as negative — these are worth a manual look." Discrepancies like these are often worth their weight in gold.

Prompts nobody thinks of first

A few prompts no one thinks of first — but which often turn out to be the most valuable. These are the "wait, that's possible?" moments: not found in any analysis handbook, but obvious in hindsight once you've seen them in action.

Detecting straight-lining:

Find respondents who answered unusually quickly or who gave the same answer throughout a Likert row. Possibly straight-lining — should we exclude them from the analysis?

A quick data-quality check you'd have to script yourself on most platforms. The AI flags suspicious cases and gives you the IDs to review.

Contradiction detector:

Find answers where the open comment contradicts the numeric score — e.g. 5 stars with a complaining comment, or 1 star with a positive comment. These are usually the most revealing cases.

A favorite of CX teams. Exactly these discrepancies often hold the real insight — either the question was understood differently, or the rating scale is being used differently than the designers assumed.

Vocabulary check:

Go through the open-ended answers from the last 6 months. Are people using our new product names or still talking about the old ones? Quote 5 examples of each.

Brand and product teams love this — it tells you whether your rename has actually landed with customers.

Suggestions for the next wave's questionnaire design:

Look at the open-ended answers. Which recurring themes come up for which I have no closed question? Suggest 3 new closed questions I should include in the next wave to capture these themes quantitatively.

The AI essentially becomes a free questionnaire-design consultant. Especially valuable for ongoing tracker studies.

Drop-off diagnosis:

My survey has a 60% completion rate. At which question do most people drop out? What distinguishes those who drop out from those who finish?

Combines case-by-case progress data with respondent profile information. It brings design problems to light (question too long, sensitive question without a preceding trust page, a technical error on a particular page) that you'd otherwise need a UX audit to find.

Translation plausibility check:

My survey runs in German and English. Compare the average satisfaction score between the two language versions. Is there a relevant difference? If so, look at the wording of question 3 in both languages — could the German version imply something different?

A common silent error in multilingual studies: a subtle wording difference in one language shifts the scores by half a point. The AI can catch this; a dashboard can't.

Combining the MCP with n8n — for automated reports

The MCP is a request-response tool: you ask, the AI answers, the conversation ends. For event-driven automation — "whenever X happens, do Y automatically" — combine the MCP with n8n.

A concrete recipe — a weekly NPS digest in Slack:

1. In n8n, set up a Schedule trigger for every Monday at 9 a.m.

2. Add an HTTP request node that calls Claude (or OpenAI) with a prompt like this: "Summarize last week's NPS responses from the QUESTIONSTAR survey code XYZ. Top 3 themes from negative comments. Format as Slack markdown."

3. The AI calls the MCP tools, fetches the data and builds the summary.

4. Add a Slack node that posts the summary to your team channel.

Every Monday morning your team gets a fresh digest, built by the AI — without anyone touching a dashboard. Other recipes follow the same pattern: new high-priority response → instant alert; monthly cohort comparison → into a Notion page; quarterly review → into an email draft.

Get started in a few steps

The QUESTIONSTAR MCP server is reachable at a single URL — the connection works the same everywhere, no matter which AI client you use.

The universal 4 steps:

1. Sign in to your QUESTIONSTAR account.
2. In your AI client, add an MCP server — with the URL https://mcp.questionstar.com/mcp and a name of your choice (e.g. “QUESTIONSTAR”).
3. Choose OAuth as the authentication method and start the process (depending on the client, the button is labelled “Authenticate”, “Connect” or “Sign in”).
4. In the QUESTIONSTAR dialog, confirm that the AI client may access your surveys and responses.

The AI client is now connected and you can get started right away.

Step-by-step guides

The exact click paths differ between clients. Detailed step-by-step guides with screenshots:

• Connect the QUESTIONSTAR MCP with Claude
• Connect the QUESTIONSTAR MCP with ChatGPT

Other MCP-compatible clients (e.g. Cursor, Zed, Kiro, Codex CLI or your own agents) follow the same pattern: add an MCP server, enter the URL, approve the OAuth connection. For the exact names of the entries in each menu, see the client’s documentation.

Your first prompt

Once your client is connected, open a new chat and try the following:

Which of my QUESTIONSTAR surveys are currently running?

The AI calls list_surveys, filters by status and shows you a clear list. The first time, your client asks for permission for the tool call — click Allow. That gives you a working MCP integration. Everything else builds on it.

Privacy, security and GDPR

Connecting your survey data to an AI client is a serious decision, not an incidental step. What actually happens — and what to watch out for.

How the data flows

• The MCP server never pushes data anywhere on its own. It sits idle until an AI client makes a tool call. Calls only happen when you ask the AI to do something that requires it.
• Data only leaves QUESTIONSTAR as part of a specific tool call. When you request an executive summary, the AI fetches the responses it needs for that — not your entire account.
• The AI client (Claude, ChatGPT, …) processes this data on its own infrastructure for the duration of your prompt. What happens after the chat depends on the client's data-retention policy — see "Choose your AI client carefully" below.

Private data of your respondents stays out

Personal details of your respondents — names, email addresses and similar contact data — are deliberately not returned by the MCP. We ask for your understanding here: it simply keeps the people behind your data better protected. If you need these details, access them directly in the QUESTIONSTAR interface — there they remain available to you as usual.

Access scope

• The OAuth connection inherits your account’s permissions — and nothing beyond them. What you can’t see in your account, the AI client can’t see either.
• The connection can’t escalate privileges — it can’t access other accounts, change billing or alter permissions.
• You can revoke access at any time — the AI client in question loses the connection immediately.

Choose your AI client carefully

You entrust the AI client with all the data the MCP returns. Anthropic (Claude), OpenAI (ChatGPT) and other reputable providers have public privacy policies and offer enterprise modes with stricter retention rules. Before you connect sensitive data, read the relevant policy:

• Anthropic (Claude): By default, conversation data is not used for training. Enterprise plans additionally offer data-residency options.
• OpenAI (ChatGPT): Behavior differs between consumer and Team/Enterprise plans. Enterprise has the strictest no-training defaults.
• Cursor, Zed, etc.: These usually route through Anthropic's or OpenAI's APIs — the policy of the underlying provider applies in each case.

GDPR considerations

QUESTIONSTAR remains your processor; the MCP changes nothing about that legally. The AI client is a separate processor — review its data processing agreement (DPA) separately if your data contains personal information within the meaning of the GDPR. For especially sensitive projects (medical, governmental, minors), use an enterprise plan from your AI provider and document the processing chain in your records of processing activities.

Troubleshooting

Connection fails or "MCP server unreachable"

• Check the exact URL: https://mcp.questionstar.com/mcp. The trailing /mcp is required.
• Restart Claude Desktop after adding the connector. Some versions only load MCP servers at startup.

"Tool call not authorized" or no response

• By default, your AI client asks for permission per tool call. If the prompt seems to hang, look for a permission dialog in the client interface and click Allow.
• If you previously declined a tool and now want to allow it again: open the connector settings in Claude Desktop and re-enable the tool in question.

The AI returns an empty response or wrong data

• Sometimes the AI calls a tool with the wrong parameter (e.g. the wrong survey code). Ask it to first list the available surveys and confirm the right one — then ask your original question again.
• For complex prompts (Tier 2 / Tier 3 analysis), ask the AI to show its work: "Walk me through your calculation step by step before you give the final result." This makes errors visible, and often the AI corrects itself.

"My numbers don't match the QUESTIONSTAR dashboard"

• Date ranges. Make sure the AI and the dashboard are filtered to the same period — the AI may default to "all time" while the dashboard shows "last 30 days."
• Disqualified responses. The dashboard sometimes excludes disqualified cases automatically; the AI may include them unless you say otherwise. Ask it: "Recalculate — excluding disqualified responses."
• Open vs. completed responses. Some dashboard tiles count only fully completed submissions; get_result_cases returns all by default. Be explicit in the prompt about what you want.

Related integrations

• n8n — for event-driven workflows (new response comes in → run AI prompt → post somewhere).
• Zapier — no-code forwarding of QUESTIONSTAR events to thousands of apps.
• REST API — direct programmatic access, without the AI layer.
• Google Data Studio & Microsoft Power BI — for predefined dashboards that refresh automatically.

The MCP shines at flexible, narrative, ad-hoc work; the others shine at structured, repeatable, dashboard-based work. Most teams end up using both.