Pipette it once. Get the robot protocol.

Labs publish thousands of protocols. Robots run a hundred.

Everything else has to be programmed, and scripting liquid handlers "remains a persistent bottleneck requiring substantial programming expertise" (bioRxiv, 2025). Opentrons robots are in labs at thousands of institutions. Show, Don't Code starts from what a lab already has: the protocol they do by hand, or the one they already published.

What happens to your video

Meta models, each doing the part it is best at. Nothing reaches the robot until the simulator has run it and you have seen it.

1. 1

   Watch

   Muse Spark 1.3

   Reads the whole video and the protocol sheet lying on the bench, then looks again at each moment liquid is drawn up to tell which tube it came from.

2. 2

   Track

   SAM 3.1

   Follows the pipette tip through the recording. Click any tube and it follows that tube too, while Muse Spark throws out the frames where it jumped to a look-alike cap.

3. 3

   Listen

   Muse Voice Transcribe

   Anything you say while you work, like “fifty microlitres of primer”, becomes timestamped evidence, biased toward the reagent names on your sheet.

4. 4

   Check

   Muse Spark Open

   A reviewer agent watches the same video, runs the Opentrons simulator itself until the volumes add up, and asks you about whatever the footage can't prove.

5. 5

   Draw

   Muse Image

   Redraws each step's own frame as an illustration for a printed bench card, and Muse Spark sends back any drawing that doesn't match its frame.

The output is code a robot runs

The models decide what happened. Plain, deterministic code decides how the robot does it, so model output never becomes executable code.

# Generated by Show, Don't Code from a video of a human performing the protocol.
requirements = {'robotType': 'Flex', 'apiLevel': '2.24'}

def run(ctx):
    reagents = ctx.load_labware('opentrons_24_tuberack_eppendorf_1.5ml_safelock_snapcap', 'C2', 'reagent tubes')
    tc = ctx.load_module('thermocyclerModuleV2')
    dest = tc.load_labware('opentrons_96_wellplate_200ul_pcr_full_skirt')
    p50 = ctx.load_instrument('flex_1channel_50', 'left', tip_racks=tips50)
    liquid4 = ctx.define_liquid('PrimeSTAR Max Premix', description='25 uL drawn per reaction', display_color='#e8590c')
    lc4 = ctx.get_liquid_class(ctx.params.lc4)  # viscous: picked from the reagent's name
    # step 1 [30-45s]: Add Sterile distilled water
    for r in range(ctx.params.reactions):  # chosen on the robot's touchscreen
        p50.pick_up_tip()
        if ctx.params.detect_liquid:
            p50.require_liquid_presence(reagents['A1'])
        p50.transfer_with_liquid_class(liquid_class=lc0, volume=20.0, source=reagents['A1'],
                                       dest=dest.wells()[0 + 1 * r], new_tip='never')
        p50.drop_tip()
    # …
    tc.execute_profile(steps=[…], repetitions=ctx.params.cycles, block_max_volume=50)

And a bench card the next person can follow

Muse Image redraws each step's own video frame as a training-manual illustration. Muse Spark compares every drawing with its frame and sends it back if anything was added, moved or written in. The words under each drawing come from the protocol, never from the image model.

The app prints the card for any run with a video. Lab video: BioVL2 (Kyoto University), CC BY 4.0.

Measured on videos it never saw

We tuned the prompts on one PCR video and scored the pipeline against human step annotations on videos it never saw from the public BioVL2 dataset, including one protocol it was never developed on.

On PCR, about one proposed step in three is extra or mislabelled; on a protocol it was never developed on, most reagents come out unnamed. That is why the app shows every step next to the footage, lets you name a tube with one click, and asks before it writes code.

Where it still needs you

A camera can't see everything a protocol needs. The app asks once, in plain words, instead of guessing.