K-Dense has released LabMCP, a free open-source toolkit that lets AI assistants like Claude operate laboratory instruments through plain-language requests. A researcher can ask the assistant to weigh a sample, heat a solution, or log sensor readings, and the AI handles the instrument commands while staying within safety limits the user sets.
The system works through the Model Context Protocol (MCP), an open standard that connects AI applications to external tools. Each instrument family gets a small connector program that translates between the AI's requests and the instrument's command language. The AI never writes raw instrument code or guesses at command syntax.
How the safety layers work
LabMCP includes several guardrails before any command reaches physical hardware. A practice mode simulates instruments so users can rehearse experiments with nothing connected. A read-only mode lets the AI observe measurements without changing settings. Users can set hard limits - such as a maximum temperature of 60°C - and any request above that threshold is refused before transmission.
Actions that heat, move, dispense, or switch on power are flagged as hazardous. The AI app asks for explicit approval before running them. Every command and reply is logged, creating a record the user can review or export. The project's documentation said the tool is "a research tool, not a certified safety system or medical device."
What instruments it supports
The initial release covers balances from Mettler Toledo and Sartorius, IKA hotplate stirrers, New Era and Tecan Cavro syringe pumps, Ocean Insight spectrometers, PalmSens potentiostats, Keithley source meters, Lake Shore temperature controllers, Pfeiffer vacuum gauges, and SRS lock-in amplifiers, among others. Universal protocol connectors handle SCPI, Modbus, EPICS, and SiLA 2 devices, covering instruments from many brands that speak those standards.
Biology and life sciences connectors include Atlas Scientific EZO sensors for pH and dissolved oxygen, Micro-Manager microscope control, Opentrons OT-2 and Flex liquid handlers, and BrainFlow biosensing boards. A Bluetooth LE health sensor connector reads heart rate, SpO2, blood pressure, and temperature data, marked for research use only.
All connectors are currently listed as simulated - built from manufacturer manuals and tested in practice mode - and await verification by users who own the real instruments. "No coding needed for the most valuable contribution: testing a connector on your real instrument," the project page said.
Setup takes minutes
Users install a single helper tool called uv, then paste a few lines into their AI app's MCP settings. Switching from practice mode to a real instrument requires changing one line in that same configuration file. A built-in port scanner identifies connected instruments, and a connection check command confirms the instrument responds before any AI interaction begins.
Multiple instruments can run simultaneously. The project provides an example configuration that sets up six practice instruments at once - a balance, stirrer, pH probe, syringe pump, potentiostat, and spectrometer - letting users test coordinated multi-instrument workflows.
Why this matters for science and research professionals
LabMCP removes the programming barrier between researchers and their instruments. A chemist who knows what experiment to run but not how to script a serial command can describe the procedure in plain language and let the AI execute it. The safety limits and command logging address the primary concern with AI-controlled hardware: that a misinterpreted request could damage equipment or create hazards.
The open-source model means instrument support grows as users contribute connectors and verify existing ones on real hardware. For labs evaluating AI integration, the practice mode provides a zero-risk way to test workflows before committing to live instrument control. Professionals looking to build these skills can explore MCP Courses or follow the AI Scientific Research Courses learning path.
Your membership also unlocks:
