Miqyas Beta
Log in

Ask your data a question in plain language

Connect Claude, ChatGPT, or another Model Context Protocol assistant to surveys you can already read. The assistant receives aggregates computed by Miqyas, never a table of individual responses. You approve the connection, see its questions in your profile, and can withdraw it in one click.

Talk to us See all features

Use the assistant your team already has

Miqyas uses the Model Context Protocol to connect an assistant to a data source, with OAuth 2.1 authorization. You do not paste a key or install another application.

Claude

Add Miqyas as a custom connector in Claude’s settings. Sign in once, approve the connection, and it appears in your tools.

ChatGPT

Add it as a connector in the same way. Approval, access records, and withdrawal work the same way as they do for Claude.

Anything else that speaks MCP

Local runtimes and self-hosted assistants use the same endpoint. This suits policies that require the model to stay on your own infrastructure.

What a query looks like

These two examples show the shape of a real exchange. The assistant does not receive your data as a file. It asks Miqyas for one aggregate, and Miqyas calculates and returns it.

These exchanges are illustrative, not recordings. The tool calls and answer shapes match the connector, but the figures are examples.

Break down a region without opening the dashboard

A field team has 1,240 household responses across Syria and wants to see where malnutrition indicators cluster before the next distribution round.

The researcher asks

In the household survey, what is the average number of meals per day by governorate, and where is coverage thinnest?

The assistant calls

listSurveys

Finds the surveys this researcher may read. Surveys outside the organization, and special-category surveys without a recorded exception, stay out of the list.

The assistant calls

listFields · surveyId: household-2026

Reads the question catalogue. Question identifiers are paths, so the next query uses fields that actually exist in the survey.

The assistant calls

runQuery · avg(/nutrition/meals_per_day) · group by region · Syria - Governorates

Miqyas matches each response to the administrative area containing its recorded location. The result includes area names, not only codes.

Miqyas returns

Across 1,240 households, the average is 2.4 meals per day. Three governorates are well below that figure. The two lowest also have the smallest samples, so the gap may be real but their order is not reliable yet.

GovernorateAverage mealsResponses
Rif Dimashq2.6412
Aleppo2.5388
Idlib2.1196
Deir ez-Zor1.961
Quneitra1.824
No location recorded2.3159

Boundaries: UN OCHA Common Operational Dataset, via HDX (CC BY 3.0 IGO). Responses without a location, or outside every area, appear in their own row instead of disappearing from the total.

The assistant never sees a coordinate or an individual household. It receives one aggregate table with a sample size for each row. That context lets it qualify the ranking instead of presenting five numbers as equally reliable.

Triage data before a deadline

A coordinator has a report due and wants to check the incoming data without opening a thousand submissions.

The researcher asks

Is anything wrong with this week’s submissions? Should I check on any collector?

The assistant calls

runQuery · count() · group by review state · submission date

Review state, fill duration, respondent country, and collector are chartable alongside survey answers, so this takes one query. There is no export step.

Miqyas returns

318 submissions arrived this week: 291 approved, 19 pending review, and 8 rejected. The rejections cluster on Tuesday. The median fill time that day is about half the study baseline, which the quality score flags as rushing.

DayApprovedRejectedMedian fill time
Monday64011m 20s
Tuesday5865m 40s
Wednesday71110m 55s
Thursday98112m 10s

The researcher asks

Which collector was working Tuesday?

Miqyas returns

Two collectors submitted on Tuesday, and almost all rejections came from one of them. That is worth checking, but it is only six submissions, so do not draw a conclusion yet.

You can ask "should I be worried?" directly. Reaching the same view in the dashboard would take four charts, while this answer includes its caveats.

How to connect

Set it up once. There is no secret to copy, store, or share.

  1. 1

    An owner enables access

    Assistant access starts off. An owner or admin can enable it for the organization and disable it again at any time. Disabling access stops existing connections as well as new ones.

  2. 2

    Add Miqyas to your assistant

    Add your Miqyas address in Claude or ChatGPT connector settings. The assistant handles the rest, so you never generate or store a secret.

  3. 3

    Approve it in Miqyas

    Miqyas opens a sign-in and approval screen that names the assistant and the address it will send data to. You make the decision inside Miqyas.

  4. 4

    Keep control of the connection

    Your profile lists connected assistants and their last use. Withdrawing one takes a click and applies to its next request. An administrator can withdraw any connection from organization settings.

What the assistant can reach

The connection uses your existing permissions. Miqyas checks each limit on every request, regardless of how the model phrases its question.

Aggregates, not individual responses

The connector can compute counts, averages, medians, and breakdowns. It cannot return an individual submission, free-text answer, or respondent identifier.

Coordinates stay on the platform

Miqyas does not return locations as points. For geographic questions, it matches each response to its administrative area and returns the areas instead.

The same access you already have

Miqyas recomputes permissions from your current role on every request. If your access changes or you leave the organization, the connection stops on its next call.

Special-category surveys stay out

Surveys that collect health, religion, or other GDPR special-category data stay excluded unless an administrator grants a recorded exception for that survey.

Small groups are withheld openly

Miqyas suppresses groups below the minimum size and reports how many it withheld. The page shows when the total no longer adds up, so a partial answer is not mistaken for a complete one.

Every access is recorded

Each request records the assistant, researcher, survey, and questions it touched. Filter values stay out of the log because a respondent’s answers are not audit data.

What this does not do

This does not hand an assistant your dataset. Nothing is uploaded, copied, or used for training. The assistant asks a question, Miqyas computes one aggregate, and only that aggregate leaves the platform.

It is not a substitute for reading the data. A model can misread a correct result, and a small sample can look like a finding without context. Miqyas returns sample sizes, suppression counts, and truncation flags so the answer includes its caveats. The judgement remains yours.

You do not have to use it. Assistant access stays off until someone enables it, and the dashboard remains available on its own. If your policy does not allow sending research questions to a third-party model, you can leave this feature off.

Try it with your own data

Assistant access is available to every Miqyas organization during the beta. Talk to us if you want a walkthrough with your surveys or if your data protection officer needs the details first.

Talk to us