A harness for AI in lab data analysis · Early release

Analyze your data with the software you already trust.

Open source under the GNU Affero General Public License (AGPL-3.0)

An agent harness for data analysis.

$curl -fsSL https://cuvette.ai/install.sh | sh

Type this command in the Terminal app on macOS or Linux. It needs Node.js 22, a free program. Then type cuvette app to open Cuvette in your browser. For Windows, see Install.

Fig. 1 | How a question moves through Cuvette. A, Your data stay in their folder. B, The harness takes the model through four steps, and you approve each decision. C, Cuvette runs programs that scientists already use. An adapter tells Cuvette how to run one program. The figure shows one program from each field. The answer and the record come out as an electronic lab notebook (.eln) file or a protocol.

How a session works

Work with your data and your analysis programs in plain language.

Ask a question in your own words. The harness turns it into steps in real programs, and you approve the method. Each session has the same four steps. Pick a field to see an example.

Questions

The model asks before it guesses.

It reads your files and your question. Then it asks a few short questions about your experiment.

Each question shows the recommended answer and the reason for it.

plate-03 · 96 images

Question 1 of 3

Which channel shows the nuclei?

Recommended: DAPI. The file names end in w1, and channel 1 has the stain.

DAPI channel 1

GFP channel 2

Both channels

bone-marrow · 2 samples

Question 1 of 4

Which cells do you keep?

Recommended: at least 200 genes and at most 10% mitochondrial reads. The paper uses these cutoffs.

200 genes, 10% mitochondrial from the paper

500 genes, 5% mitochondrial stricter

Keep all cells

mouse-a2 · 3 sections

Question 2 of 3

How do you normalize each pixel?

Recommended: TIC, the total ion current. It is the default in the vignette.

TIC total ion current

RMS root mean square

No normalization

Magnetoencephalography (MEG) · 1 recording

Question 3 of 4

Which time window do you measure?

Recommended: 80 to 120 ms after the tone. The N100 peak falls in this window.

80 to 120 ms around the N100

50 to 150 ms wider

The whole epoch

Proposed metrics

You see the method in plain words.

The model writes what it will measure and how. Each setting that can change a result is on the list.

You change any line, or you approve the list.

Proposed metrics, version 1

Nucleus count for each imageObjects in the DAPI channel

Threshold: Otsu, globalOne threshold for each image

Remove objects under 30 pixelsDebris and noise

Keep nuclei at the image edgeYour answer to question 3

Change a lineShow me on a sample

Proposed metrics, version 1

Remove cells with fewer than 200 genesEmpty droplets and broken cells

Remove cells with more than 10% mitochondrial readsDying cells

Normalize each cell, then log transformSame total count for each cell

Clusters and their marker genesNeighbor graph, then Leiden clusters

Change a lineShow me on a sample

Proposed metrics, version 1

Normalize each pixel to its TICTotal ion current

Pick peaks, then align them across pixelsOne peak list for the section

Segment the tissue into regionsSpatial shrunken centroids

Mean intensity of each peak in each regionOne table for each section

Change a lineShow me on a sample

Proposed metrics, version 1

Epochs from 0.2 s before to 0.5 s after each toneOne epoch for each event

Baseline: from the start of the epoch to the toneThe tutorial default

Leave out the channels that the file marks as badTwo channels

Peak amplitude in the N100 windowFor each condition

Change a lineShow me on a sample

Demonstration on a sample

It runs on one sample first.

The method runs on one sample only. You look at the result before the other samples run.

If the result looks wrong, try another sample or change the metrics.

Demonstration on a sample · A01_s1

24 nuclei in the drawing

Confirm and run allTry another sample

Demonstration on a sample · pulse

Green dots: 393 of 510 cells kept in the drawing

Confirm and run allTry another sample

Demonstration on a sample · section s1

3 regions in the drawing

Confirm and run allTry another sample

Demonstration on a sample · left auditory

Shaded: the N100 window, 80–120 ms

Confirm and run allTry another sample

Confirm

Then the batch runs, and the record stays.

After you confirm, the method runs on all samples. Each number in the answer must match a result in the session record.

Export the session as an electronic lab notebook (.eln) file, or as a protocol for a lab member.

Confirmed version 1

Running on 96 images

Method confirmed by you
Each step logged with its inputs
Each number checked against the record
Export: .eln notebook file or protocol

Confirmed version 1

Running on 2 samples

Method confirmed by you
Each step logged with its inputs
Each number checked against the record
Export: .eln notebook file or protocol

Confirmed version 1

Running on 3 sections

Method confirmed by you
Each step logged with its inputs
Each number checked against the record
Export: .eln notebook file or protocol

Confirmed version 1

Running on 4 conditions

Method confirmed by you
Each step logged with its inputs
Each number checked against the record
Export: .eln notebook file or protocol

Fig. 2 | The four steps of a session. a, Questions. b, Proposed metrics. c, Demonstration on a sample. d, Confirm. The panels are drawings of the app screens. Each example uses the decisions of a case from the validation set. The data in panel c are drawings, not results.

Programs

One harness. Your choice of AI model. 61 programs and growing.

Use Claude from Anthropic, a model that runs on your own computer, or any model server that uses the OpenAI format. An adapter connects the harness to one program. Pick a field to see its programs. Other people can write and share adapters: see community adapters.

Plate I | Imaging. Count and measure cells, nuclei and tissue in microscope images.

See the imaging adapters

Records

Data provenance, built for peer review.

A reviewer can trace each number back to the step and the file that made it. These rules are code that runs on each step. They are not instructions to the model.

You make the decisions

Each adapter lists the choices that can change a result. If a step needs a choice that you did not make, the step stops. The harness compares the options, then asks you.

You can repeat each step by hand

Each step shows the command or the menu path with its dialog values. Type /export protocol to write the steps as a protocol.

Every step goes in one record

Cuvette adds to the session record and never deletes from it. Each result keeps its settings and the fingerprints of its input and output files. If you change a decision, Cuvette marks the old results as out of date.

Each number has a source

Each number in the answer must match a result in the session record. A second model reads only the session record and checks the answer. The model must correct each error before you see it.

Validation

We repeat published analyses and compare the numbers.

The model gets the data and the question from a paper or an official tutorial. Then we compare its answers with the known values.

published analyses: 33 research papers and 20 tutorials or software test sets

472

known values that we score

Claude models in the final run: Opus, Sonnet and Haiku, three blind runs each

Where the 472 known values come from

Printed in the paper	276	58%
Printed in an official tutorial	51	11%
Calculated by us with a different program	133	28%
Calculated by us with the same program	12	3%

The chart shows sources, not scores.

The source of each value is in the sources document. See the validation method and each paper.

Install

Two ways to start.

The terminal is the window where you type commands. Use it if Node.js is on your computer. Node.js is a free program that runs Cuvette. If you do not have it, use the desktop app.

Terminal

For macOS, Linux and Windows. Cuvette needs Node.js 22 or later.

Install Cuvette. npm, the installer that comes with Node.js, downloads it.
```
npm i -g @elparko/cuvette
```
Check your computer. This command finds the programs and models that Cuvette can use.
```
cuvette doctor
```
Open the app in your browser.
```
cuvette app
```

Desktop app

For macOS. One .dmg file, the Mac installer. You do not need the terminal.

Download the .dmg file and drag Cuvette to Applications.
Open it the first time. Apple has not signed (approved) the app yet, so macOS blocks it. Open System Settings, then Privacy & Security. Click Open Anyway.
Use it from Finder. Right-click a data folder and choose Analyze with Cuvette.

Download for Mac

The command was ga before we renamed it cuvette. Until the new name is in the released version, type node bin/ga.js in the source folder. See all install methods.

Ask a question about your data. Keep every decision.

Install Cuvette Source on GitHub