Lab robots get tested at one setting. Tare checks all of them.
Tare reads the program a liquid-handling robot will run and proves it is safe for every value it allows, before the robot moves. When it isn't, Tare names the value and the line.
The settings nobody tried are where runs fail.
A protocol that allows anything from 10 to 1000 µL is usually checked once, at its default. A scientist, or an optimiser running experiments overnight, then picks a value no one tested. A tip overfills, a well runs dry, a sample spills into its neighbour.
Tare treats each setting as an unknown and checks the whole allowed range at once, without running the code. It covers overflowing and empty wells, tip volumes, running out of tips, crashes, reagent supply, cross-contamination and module state.
It works on Opentrons OT-2 and Flex protocols today, and on PyLabRobot programs, which drive Hamilton, Tecan and other robots.
An agent can check its own plan before it reaches the robot.
Tare runs as an MCP server, so an AI agent can call it as a tool. Given a buggy public protocol and Tare as its only checker, Claude Code found the bug, fixed it, flagged the one change a person should confirm, and wrote a run record. It took five minutes and cost $0.89. An independent simulator check passed the fix at all 12 test settings.
Early results on real protocols.
Protocols that Opentrons ships for its Flex robot crash at settings they allow, while their defaults pass. Shake speeds below the heater-shaker's minimum, volumes that overfill a tip, a tip layout the pipette can't use. Each failing step, reproduced in Opentrons' own analysis tool, failed exactly as Tare predicted: 20 of 20 checks.
Public Opentrons protocols with a bug that only shows up away from the defaults. One bug per protocol was replayed in the simulator: 62 of 62 confirmed.
Buggy protocols an AI fixed with Tare's feedback, against the standard simulator's. 50 tasks, 600 attempts, one model, graded independently. Sampling the simulator at eight settings did about as well (66%).
One Tare check, against a median of about 20 random simulator runs to find the bugs an eight-run sample missed. The worst case took 131.
Next: a camera that confirms what actually landed in each well, now in prototype, and tests on real robot runs.
Try it on your protocols.
We're working with our first labs. Send two or three protocols and we'll send back what Tare finds.