LGI — LAN GPIB Inventory

Finds Agilent/Keysight E5810A-class LAN/GPIB gateways on the network, walks the GPIB bus behind each one, and keeps a SQLite record of every instrument it has ever seen — gateway, GPIB address, instrument identity, and first/last seen.

Standard library only. No VISA runtime, no pyvisa, no python-vxi11. Python 3.9+ with Tk (python3-tk on Debian/Ubuntu, bundled on macOS/Windows).

python3 lgi.py                       # default DB at ~/.lgi/inventory.sqlite3
python3 lgi.py --db bench.sqlite3

Files

File	Purpose
lgi.py	Tkinter GUI. Gateways tab, one tab per gateway, inventory browser.
lgi_core.py	ONC RPC + VXI-11 client, discovery, scanning, SQLite layer, headless CLI.
lgi_testcontroller.py	Optional: reads TestController device definitions and matches them to the inventory.
fake_gateway.py	A simulated E5810A with four virtual instruments, for testing without hardware.
lgi.spec	PyInstaller build: one binary per platform, .app bundle on macOS.
build.yml	GitHub Actions workflow for Linux, Windows and both macOS architectures.

How discovery works

Two independent mechanisms, both selectable:

1. Broadcast — a portmap GETPORT for the VXI-11 core program (0x0607AF, version 1) sent to UDP/111 on each subnet broadcast address. Every VXI-11 instrument server answers with its core channel port. Same mechanism VISA uses for TCPIP?*::INSTR.
2. Subnet sweep — a threaded unicast GETPORT over TCP/111 across a CIDR. Slower, but crosses routers and survives switches that drop directed broadcasts. Capped at 4096 addresses.

Responders are then identified over HTTP: /lxi/identification first (E5810B and other LXI devices give clean XML), falling back to scraping the welcome page for model, serial, firmware and hostname. Anything that answers on the VXI-11 program is listed even if identification fails — the tool works with E2050s and clones, not just the E5810A.

If neither finds your gateway, Add by address… takes an IP or hostname, and optionally an explicit VXI-11 core port to bypass the portmapper entirely (10.0.0.50:1024 also works in the address field).

How the bus scan works

Two-phase, and it does not disturb instrument state:

1. Serial poll every address. Create a link to gpib0,<addr>, device_readstb with a short timeout, destroy the link. Fast, and an empty address just times out at the bus level.
2. *IDN? the responders. Write, read, parse into manufacturer / model / serial / firmware.

The status byte is decoded (RQS/MSS, ESB, MAV, QSB, EAV) and kept, so an instrument sitting with SRQ asserted is visible at a glance.

Options worth knowing:

• Also query addresses that ignore serial poll — catches older listen-mostly gear that never answers a poll but does answer *IDN?. Slower: every empty address costs a full *IDN? timeout.
• Send device clear before *IDN? — off by default, since DCL will abort whatever the instrument is doing. Useful for something wedged mid-transfer.
• Controller at / skip it — the E5810A's own GPIB address, default 21.
• Interface — gpib0 for the E5810A. Editable for gateways using other logical names.

Anything the gateway reports as present gets a row; addresses that answer serial poll but stay mute to *IDN? are recorded too, marked "present (mute)", so you can annotate them by hand.

Console

Per-gateway terminal against a single address. A link stays open until you change address or drop it, so multi-step sequences work. A command ending in ? is written then read back; anything else is write-only. Up/Down walks history. Buttons for serial poll, bare read, device clear, and go-to-local.

TestController integration (optional)

Off by default, and it lives on its own sub-tab so it stays out of the way when unused: Inventory → TestController. Turn it on there, point it at your TestController install or working folder, and press Scan definitions.

That sub-tab shows two tables side by side — your inventoried instruments with the definition file matched to each, and every definition the scan found, with its handle and the ports it declares. A log below records what was read. Over on the Instruments sub-tab, two extra columns appear on the records themselves so the match is visible alongside everything else known about the instrument. Nothing is ever written to the TestController installation.

Matching is on #idString, which is the first two comma-separated fields of the instrument's *idn? reply, and which TestController's SCPI driver requires to match before it will connect. Since the inventory already stores the full IDN read off the bus, both sides of the comparison came from the instrument itself rather than from a guess about model names.

Match	Meaning
exact	#idString matches the recorded IDN. TestController would accept this.
case differs	Matches only when case is ignored — worth checking, as the comparison is literal.
likely / possible	No IDN to compare, so the model name was matched against #name or the file name. Confirm it yourself.
none	No definition in the catalogue mentions this instrument.

An unambiguous #idString hit is linked automatically. Anything ambiguous — two files claiming the same instrument, or only a name-based guess — is left alone and marked, because a wrong link is worse than no link. Select the row and press Choose driver… (or double-click it) to see every candidate with the reason it matched, and pick one. A choice you make by hand is marked "chosen" and survives later rescans; automatic links are recalculated each time.

Case handling, which is the fiddly part:

• The folder is found whether it is spelled Devices, devices or DEVICES, and both the install folder and the working folder under Documents/TestController are searched.
• File names are matched case-insensitively, so Keithley2001.txt and keithley2001.txt are the same definition as far as matching goes.
• Two files whose names differ only by case are reported as a collision. They coexist happily on Linux but collapse into one file on Windows and macOS, so a definition set carrying such a pair silently loses one when it moves between machines.

A #meta block is a template rather than a device, so it is never offered. Devices built from it with #metadef are, inheriting any tag they don't override. If a matched definition's #port doesn't list GPIB, that's flagged — TestController won't offer it for a GPIB connection as written.

Database

gateways → instruments, plus scan_runs / scan_hits for scan history and diffing, tc_links for TestController associations, and app_settings for preferences. Timestamps are ISO-8601 UTC in the file, shown as local time.

Instruments are keyed on (gateway_id, gpib_address, idn). Swap the instrument at address 16 and you get a second row rather than a clobbered one — the old instrument keeps its last_seen, so the record shows what was on that address and when. first_seen never moves once set.

nickname and notes are yours to fill in and survive rescans. Export is JSON or CSV, per-bus or whole-database.

Testing without hardware

sudo python3 fake_gateway.py                     # portmapper on 111, web on 80
python3 lgi.py                                   # sweep 127.0.0.1/32

Or without root:

python3 fake_gateway.py --pmap-port 11111 --core-port 9010 --no-http

then add 127.0.0.1:9010 by address. The simulator presents a Keithley 2001 at 2, a 2002 at 9, a TDS 460A at 16, and a deliberately mute device at 23.

Building binaries

pip install pyinstaller
pyinstaller lgi.spec --noconfirm

Platform	Output	Notes
Linux	dist/lgi	Build on the oldest glibc you support; python3-tk must be installed.
Windows	dist/lgi.exe, dist/lgi-cli.exe	A windowed exe has no stdout, so the subcommands get a console build.
macOS	dist/LGI.app, dist/lgi	Build against python.org Python — the system one links Tk 8.5.

The GUI entry point forwards subcommands to the headless side, so one binary does both: lgi opens the window, lgi scan 192.168.1.50 does not. --db works on either side of the subcommand.

Three platform quirks worth knowing before you ship a build:

• Windows Firewall eats broadcast replies. The replies to a subnet broadcast come from many different addresses and do not match the outbound UDP mapping, so they are dropped unless a rule allows them. The subnet sweep is unaffected and finds the same gateways, so this degrades rather than breaks — but the first run on Windows may look like nothing is there.
• macOS needs local network permission. From macOS 14, an app is refused local network access unless its Info.plist carries NSLocalNetworkUsageDescription; the spec sets it. The prompt appears on first discovery, and denying it makes both broadcast and sweep return nothing.
• Gatekeeper. The workflow ad-hoc signs the bundle, which is enough to run it on the machine that built it. Anything you hand to someone else needs a Developer ID identity and notarisation, or they get "cannot be opened because the developer cannot be verified".

Headless

The core module runs without Tk:

python3 lgi_core.py discover --cidr 192.168.1.0/24
python3 lgi_core.py scan 192.168.1.50 --deep
python3 lgi_core.py list
python3 lgi_core.py export --out inventory.json

python3 lgi_core.py tc --base ~/TestController          # report matches
python3 lgi_core.py tc --link                           # and record them

Same database, so a cron scan and the GUI share one inventory.

The gateway's own settings are not touched by this tool — configure the E5810A from its web interface.

Notes on E5810A behaviour

• Link creation usually succeeds regardless of whether a device is actually at that address; the failure surfaces at the first I/O as error 15 (I/O timeout). That is why presence detection is serial-poll based rather than link-based.
• Concurrent links are limited. Every scan link is destroyed immediately after use, and the console link is dropped before a scan starts.
• The core channel occasionally drops on older firmware. The scanner reconnects once per address and carries on rather than aborting the pass.
• Scan time is roughly addresses × poll timeout for the empty ones. 500 ms across 0–30 is about 15 s worst case; drop to 200–300 ms on a healthy bus.