Control Lab Instruments with ChatGPT or Claude in 2026 (3 Ways)

You can control lab instruments with ChatGPT or Claude in three ways: have the model write a PyVISA script that you run, connect the model to the instruments through an MCP server, or use an AI agent built for the bench such as TestFlow (testflowinc.com). Neither ChatGPT nor Claude can reach a USB, LAN, or GPIB instrument from a browser tab, so every route puts some software on the bench PC between the model and the hardware.

This guide is for engineers who have already tried pasting a programming guide into a chat. It shows correct scripts for a Keysight E36313A power supply and a Keysight InfiniiVision oscilloscope, the five ways LLM-written instrument code goes wrong, the MCP servers that exist as of October 2026, and a SCPI safety checklist that applies to all three routes.

The 3 ways at a glance

The three ways differ in who sends each command to the instrument: your script, the model through an MCP server, or a workflow an agent built.

Claude Pro includes Claude Code, which matters for way 2, because Claude Code and Claude Desktop are MCP clients. The full product comparison is in AI test automation for lab instruments.

Way 1: ChatGPT or Claude writes a PyVISA script you run

Way 1 is the simplest route: the model drafts Python, you read it, and you run it on a PC with PyVISA and a VISA library such as Keysight IO Libraries. The quality of the script depends almost entirely on what you put in the prompt.

Give the model the exact model number, the interface, the set points, the limits, and the relevant pages of the programming guide. A prompt that works:

Write a PyVISA script for a Keysight E36313A triple-output supply on LAN at
192.168.1.50. Output 1 only: 3.3 V, current limit 200 mA, over-voltage trip
3.6 V, over-current protection on. Enable the output, wait 0.5 s, read
voltage and current, then turn the output off in a finally block. Use the
(@1) channel-list form from the E36300 Series programming guide and check
SYST:ERR? after configuration.

A correct result for that prompt, checked against the Keysight E36300 Series programming guide:

import time
import pyvisa

PSU = "TCPIP0::192.168.1.50::inst0::INSTR"  # Keysight E36313A, confirm with rm.list_resources()


def drain_errors(inst):
    """Read SYST:ERR? until the queue reports no error. SCPI never raises."""
    errors = []
    while True:
        reply = inst.query("SYST:ERR?").strip()
        if reply.startswith(("+0,", "0,")):
            return errors
        errors.append(reply)
        if len(errors) > 20:
            return errors


rm = pyvisa.ResourceManager()
psu = rm.open_resource(PSU)
psu.timeout = 5000
psu.read_termination = "\n"

print(psu.query("*IDN?").strip())
psu.write("*RST")   # outputs off, voltage at minimum, current limit at default (10 A on output 1)
psu.write("*CLS")   # *RST does not clear the error queue on this supply
psu.query("*OPC?")

try:
    psu.write("CURR 0.2,(@1)")           # compliance limit first
    psu.write("VOLT:PROT 3.6,(@1)")      # over-voltage trip level
    psu.write("CURR:PROT:STAT ON,(@1)")  # over-current protection
    psu.write("VOLT 3.3,(@1)")
    errors = drain_errors(psu)
    if errors:
        raise RuntimeError(f"setup rejected: {errors}")

    psu.write("OUTP ON,(@1)")
    time.sleep(0.5)
    volts = float(psu.query("MEAS:VOLT? (@1)"))
    amps = float(psu.query("MEAS:CURR? (@1)"))
    print(f"Output 1: {volts:.4f} V, {amps * 1000:.2f} mA")
finally:
    psu.write("OUTP OFF,(@1)")
    psu.close()

Every command above is in the E36300 Series guide. The (@1) channel list matters, because on this supply VOLT without one programs whichever output INSTrument:SELect last chose. The Keysight E36313A programming guide covers the rest of the command set, and the PyVISA tutorial covers sessions, timeouts, and backends.

Controlling an oscilloscope with ChatGPT

Controlling an oscilloscope with ChatGPT works the same way, with one extra trap: a scope returns a number even when it could not make the measurement. The script below is for a Keysight InfiniiVision DSOX2004A and uses commands from the InfiniiVision 2000 X-Series programmer's guide.

import pyvisa

rm = pyvisa.ResourceManager()
scope = rm.open_resource("TCPIP0::192.168.1.60::inst0::INSTR")  # InfiniiVision DSOX2004A
scope.timeout = 10000

scope.write("*RST")
scope.write("*CLS")
scope.write(":CHANnel1:PROBe 10")
scope.write(":CHANnel1:SCALe 0.5")            # volts per division
scope.write(":TIMebase:SCALe 200E-6")         # seconds per division
scope.write(":TRIGger:EDGE:SOURce CHANnel1")
scope.write(":TRIGger:EDGE:LEVel 1.0")
scope.write(":DIGitize CHANnel1")             # one acquisition, then the scope stops
scope.query("*OPC?")

vpp = float(scope.query(":MEASure:VPP? CHANnel1"))
freq = float(scope.query(":MEASure:FREQuency? CHANnel1"))
if vpp > 9e37 or freq > 9e37:
    raise RuntimeError("measurement could not be made, scope returned 9.9E+37")

print(f"Vpp {vpp:.3f} V, frequency {freq:.1f} Hz")
print(scope.query(":SYSTem:ERRor?").strip())
scope.close()

The 9.9E+37 check is the line a model most often leaves out. Keysight's guide says that value is returned when a measurement cannot be made, typically because the right part of the waveform is not on screen. A script without the check logs 99 undecillion volts as a reading. For Tektronix and Rigol scopes, see automating oscilloscope measurements.

Five ways an LLM-written instrument script goes wrong

LLM-written instrument scripts fail in five recurring ways, and all five pass a syntax check and run without a Python exception.

1. 1

   Hallucinated SCPI. The model borrows a command from another vendor or a sibling model. A Tektronix-style MEASUrement:IMMed:TYPe PK2PK sent to an InfiniiVision scope is not in its tree, so the scope queues -113,"Undefined header" and the script carries on. Sibling models differ too: the E36300 guide marks the MMEMory data-log export as supported only on the E36312A and E36313A.

2. 2

   Wrong subsystem or channel. On a triple-output supply, VOLT 3.3 without a channel list goes to the output currently selected, which may be the 25 V output, not the 6 V one you meant.

3. 3

   Output left enabled. OUTP ON followed by an exception, with no finally, leaves the device under test powered until someone walks over.

4. 4

   No compliance limit. After *RST, output 1 of an E36313A has a 10 A default current limit and its over-voltage trip at the 6.6 V maximum. A script that sets only a voltage protects nothing.

5. 5

   A sentinel read as a measurement. The 9.9E+37 value above, or a reading taken before the output settled, goes straight into the log.

Each failure maps to a line in the checklist further down. The habit that catches the first two is draining SYST:ERR? after every configuration block, covered in the SCPI command cheat sheet.

Way 2: an MCP server exposes the instruments to the model

Way 2 uses the Model Context Protocol (MCP), an open standard that lets an AI client call tools: an MCP server on the bench PC exposes instrument tools such as discover, query, and write, and the model calls them during the conversation. The servers below exist as of October 2026, and their guards differ more than their features.

A Keysight session shows what the model actually does through a server. Per Keysight's tool reference, the client calls discover_instruments to scan the VISA buses, resolve_commands to map each intent to SCPI with a high, medium, or low confidence, preview_commands to explain the sequence, and execute_commands only after the engineer approves the preview. get_all_errors reads and clears the error queue with explanations. A low-confidence mapping is flagged before anything is sent, which is exactly the check a pasted chat script never gets.

Keysight's getting-started page says it was tested primarily with Claude Code, Claude Desktop, and GitHub Copilot, on Windows 10 or 11. LabMCP's README marks all 32 connectors as simulated, built from the manuals and not yet confirmed on real hardware by a user, so start in practice mode. In Claude Desktop, under Settings, Developer, Edit Config:

{
  "mcpServers": {
    "scpi": { "command": "uvx", "args": ["labmcp-scpi", "--simulate"] }
  }
}

Or in Claude Code, from a terminal:

claude mcp add scpi -- uvx labmcp-scpi --simulate

What MCP does not give you is a test. The model chooses each command in the conversation, so the same request can produce a different sequence on a different day, and none of their documentation describes a test report with limits and a verdict. Keysight's release notes list reusable saved command sequences, which helps with repeatability on Keysight hardware.

Way 3: an AI agent built for the bench

Way 3 is an AI agent built for the bench: the agent turns a described test into a saved workflow with limits, a desktop app runs that workflow on the instruments, and the report comes from the results of the run. The difference from way 2 is that the output is a test you can rerun, review, and hand to someone else, not a transcript.

TestFlow is the example used here. The agent writes limits, never measurements, and it will not widen a limit to make a test pass, as set out on the trust page. The generated code is a standalone Python package you own, visible in the code view, so you can read every SCPI string before a run. The agent works in four stages, connect, describe, build, and run and report, and the build stage produces a workflow with limits rather than a chat transcript, as shown in how TestFlow works. Instruments from different vendors can share one workflow. How that compares with writing the PyVISA yourself is covered in TestFlow vs Python and PyVISA.

Which tool to use, ranked

The ranking below favours guards on a powered bench first, then instrument reach, then whether the tool produces a rerunnable test.

1. Keysight MCP Server for Instrument Control

Free, beta, Windows only, and Keysight only. Commands come only from instrument definition files, every sequence needs engineer approval, and the tool reference requires preview_commands before execute_commands. The best choice for an all-Keysight bench driven from Claude Code or GitHub Copilot.

2. TestFlow

TestFlow is an AI agent for electronics testing: you describe the test, the agent builds a reviewable workflow with Python and SCPI for your instrument models, and the Windows desktop app runs it and reports from measured results. Free to start, then $39 or $149 a month. No DAQ cards, no macOS or Linux bench runs.

It covers instruments from Keysight, Tektronix, Keithley, Rohde & Schwarz, Siglent, Chroma, and other catalog vendors, and the agent adds most SCPI instruments outside the catalog, Rigol included, by drafting the command set from the vendor's programming manual. Engineer is $39 a month with 50 runs on real instruments, per the pricing page as of October 2026. The desktop app runs the workflow on the bench PC.

Best for

Teams that want automation without writing VIs or drivers

Cost

Free to start. Engineer $39 a month, Lab $149 a month

Tradeoff

Runs on your bench through the Windows desktop app

Try TestFlow freeBook a demo

3. Rohde & Schwarz MXO AI Connect

Free, for MXO 3, 4, 5, and 5C oscilloscopes, on Windows or macOS, with no internet connection needed. The direct answer to "control an oscilloscope with Claude" if the scope is an MXO. Approval before an action is left to your AI client.

4. LabMCP by K-Dense

Open source under Apache 2.0, with 32 connectors and a generic SCPI server that supports command allow and deny lists, read-only mode, and limits checked before anything reaches the instrument. Every connector is still marked simulated, so treat a first hardware run as a test of the connector.

5. ChatGPT or Claude with PyVISA

The cheapest and most flexible route, and the one with no guard except you. Use the prompt pattern and the checklist on this page, and keep the script in version control once it works.

SCPI safety checklist before an LLM touches a bench

A SCPI safety checklist turns the five failures above into commands you can check line by line, whichever of the three ways you use.

1. 1

   Confirm the model. *IDN? must return the exact model the script or server was written for.

2. 2

   Reset and clear. *RST then *CLS. On the E36300 Series, *RST does not clear the error queue, *CLS does.

3. 3

   Set the compliance limit before the output. CURR 0.2,(@1) before OUTP ON,(@1), never after.

4. 4

   Arm the protection. VOLT:PROT 3.6,(@1) and CURR:PROT:STAT ON,(@1) on the E36313A.

5. 5

   Drain the error queue. SYST:ERR? until it returns +0,"No error", after every configuration block.

6. 6

   Wait for completion. *OPC? after slow operations such as :DIGitize.

7. 7

   Reject sentinels. Treat 9.9E+37 from a scope as a failed measurement.

8. 8

   Turn it off in a finally block. OUTP OFF,(@1) runs whether the script succeeds or fails.

9. 9

   Simulate first. LabMCP's --simulate, Keysight's mock mode, or TestFlow's browser simulation, before any hardware.

The same checklist as the SCPI sequence it produces on an E36313A:

*IDN?
*RST
*CLS
CURR 0.2,(@1)
VOLT:PROT 3.6,(@1)
CURR:PROT:STAT ON,(@1)
VOLT 3.3,(@1)
SYST:ERR?
OUTP ON,(@1)
MEAS:VOLT? (@1)
MEAS:CURR? (@1)
OUTP OFF,(@1)
SYST:ERR?

For a full power-rail test built this way, see automating power supply testing. For the connection layer underneath every route, see VISA and SCPI instrument control.

Related reading

• •AI test automation for lab instruments, six options compared
• •Best AI tools for test engineers
• •Automate a Keysight power supply with Python
• •Python instrument control libraries compared
• •GPIB vs USB vs LAN
• •Automate Rigol and Siglent instruments without coding
• •Keysight Command Expert alternatives, for finding the right SCPI
• •Supported instruments
• •Video tutorials