What is LabMCP?
LabMCP, from K-Dense, is a collection of 32 MCP servers, each a separate Python package named labmcp-<instrument>, that let an AI client operate laboratory hardware over USB, RS-232, GPIB, Ethernet or Bluetooth. The catalogue covers Mettler Toledo and Sartorius balances, IKA hotplate stirrers, New Era and Tecan Cavro syringe pumps, Ocean Insight spectrometers, PalmSens potentiostats, Keithley source meters, Lake Shore temperature controllers and an Opentrons OT-2 and Flex bridge, plus generic SCPI, Modbus, EPICS and SiLA 2 servers that reach many more instruments. Together they register 411 tools.
The repository is Apache-2.0, written for Python 3.10 or newer on FastMCP, and has about 31 stars. The shared core, labmcp, is at version 0.1.2 on PyPI. Every server has a --simulate mode, a --read-only mode, per-experiment --limit settings and a log of every command and reply that can be saved to a JSONL file.
How to install LabMCP
uvx labmcp-mettler-toledo --simulateIs LabMCP worth it?
LabMCP takes the risk of putting an agent in charge of physical equipment more seriously than most hardware servers, and every one of its 32 servers is still unproven on real hardware. The README marks all of them as simulated: written from the manufacturer's manual and tested against a simulator, with none yet confirmed by a user on the instrument itself. That label matters more than the 411-tool count when a tool sets a hotplate to 60 °C or drives a syringe plunger.
The safety model is the reason to take it seriously. All servers share one core, so the same switches work everywhere: --simulate answers like the instrument with nothing attached, --read-only lets the agent measure but not change settings while stop and switch-off still work, and --limit values such as max_temperature_c=60 are checked before a command leaves the computer. Tools that heat, move, dispense or switch on power carry MCP hazard hints so the client asks first, and the README tells users not to grant always-allow for them. The project calls its limits an extra layer on top of the instrument's own interlocks, not a certified safety system, and says it is not validated for GxP or clinical use.
These servers control physical devices, so the network options need reading. The default is stdio on the computer cabled to the instrument. An optional HTTP transport exists for shared instruments, and the client guide says plainly that it has no authentication: bind it to 127.0.0.1 and reach it through an SSH tunnel or VPN, never the open internet. The ASTM receiver for clinical analysers can also listen on a TCP port; it redacts patient names and dates of birth by default and is labelled research use only.
The project is very young. The repository was created on 25 September 2026 and all 14 of its commits landed that day and the next, ending with labmcp 0.1.2; nothing has been pushed since. Breadth comes partly from standards: the SCPI server takes command allow and deny lists, and the Modbus server works from a register map with per-point limits, which is how instruments outside the named families get reached. Keysight and Rohde & Schwarz instruments are left to those makers' own servers by design.
Where it earns its place
A postdoc in a university electrochemistry group
Run a scan-rate series of cyclic voltammograms and check how the peak currents scale
Rehearse the series on the PalmSens server with --simulate, then point it at the potentiostat with conservative limits and let the agent step through the scan rates and export CSVs. The audit log becomes the notebook record of every command sent.
A lab manager in a small synthesis lab
Record balance and pH readings without copying them by hand
Run the Mettler Toledo and Atlas EZO servers with --read-only, so the agent can read but change nothing, and save --audit-log run.jsonl for each experiment. Free to run; the only cost is the AI client.
A physics lecturer preparing an undergraduate lab
Let students practise instrument control before they touch the cryostat
Give students the Lake Shore and SRS lock-in servers in --simulate mode, which reply like the real units, so mistakes cost nothing. Switch to the real address only once the procedure has been agreed.
Before you commit
- Setup
- Install uv, add one entry per instrument to the client config (for example uvx labmcp-mettler-toledo --simulate), restart the client and ask what is connected; the README puts this at about 10 minutes. Real hardware needs the port from uvx labmcp ports, a --check run, then --address in place of --simulate.
- Integrations
- Claude Desktop, Claude Code, Cursor, VS Code, Codex and Windsurf. Interfaces include RS-232, USB virtual COM, USBTMC, GPIB through VISA, Ethernet and Bluetooth LE, plus generic SCPI, Modbus TCP and RTU, EPICS Channel Access and SiLA 2 servers.
- Data and privacy
- Runs on the computer attached to the instrument and talks to it directly; results go to your AI client's model provider. Every command and reply is logged, optionally to a JSONL file. The HTTP transport has no built-in authentication and should stay on 127.0.0.1 behind an SSH tunnel or VPN.
- Cost at scale
- Free under Apache-2.0. The real cost is the AI client's subscription and the time to verify each server on your own instrument before trusting it.
- Limits
- All 32 servers are marked simulated, not hardware-verified. Not a certified safety system or medical device and not validated for GxP or clinical use; the instrument's own interlocks and cut-offs must stay on.
What’s good
- 32 servers and 411 tools behind one safety model
- Simulator, read-only mode and pre-send limits on every server
- Hazardous tools flagged so clients ask before running them
- SCPI, Modbus, EPICS and SiLA 2 servers reach many more instruments
What’s not
- No server yet verified on real hardware
- HTTP transport has no authentication
- All commits landed in two days in September 2026
- Not validated for GxP or clinical work
Best for
Research and teaching labs that want an agent to log readings and run routine steps on serial or networked instruments, starting in simulation and read-only mode.
LabMCP alternatives
LabMCP: common questions
- Is LabMCP worth it?
- A careful safety design for letting agents run lab equipment, held back by the fact that none of its 32 servers has yet been confirmed on a real instrument. We rate it 3.5 out of 5.
- How much does LabMCP cost?
- Free and open source (Apache-2.0). You need your own instruments and an MCP client.
- What are the downsides of LabMCP?
- No server yet verified on real hardware; hTTP transport has no authentication; all commits landed in two days in September 2026; not validated for GxP or clinical work.
- Who is LabMCP best for?
- Research and teaching labs that want an agent to log readings and run routine steps on serial or networked instruments, starting in simulation and read-only mode.
- How much does LabMCP cost for a team or in production?
- Free under Apache-2.0. The real cost is the AI client's subscription and the time to verify each server on your own instrument before trusting it.
- Is LabMCP safe for company data?
- Runs on the computer attached to the instrument and talks to it directly; results go to your AI client's model provider. Every command and reply is logged, optionally to a JSONL file. The HTTP transport has no built-in authentication and should stay on 127.0.0.1 behind an SSH tunnel or VPN.
- What are the best LabMCP alternatives?
- The closest alternatives we have reviewed are ThingsBoard MCP Server and AetherLink SDR MCP.
Answers checked Nov 14, 2026.