Verify PCB connectivity against declared engineering intent, using KiCad as the oracle.
Status: pre-alpha. All six commands work and are dogfooded in this repo's own CI. The stable surface is the report and the exit codes; the Python API will move. The design lives in
docs/DECISIONS.md.
pip install kicad-netspecNeeds KiCad 8 or newer for kicad-cli — found on PATH, via Flatpak, in a macOS bundle
or a Windows install. netspec doctor says which. The core itself has no runtime
dependencies.
Every tool that edits a KiCad design reports on its own arithmetic. Only KiCad knows what
a design actually is. netspec asks it — after every change — and fails loudly when
intent and reality diverge.
netspec doctor # find and probe a KiCad engine
netspec netlist board.kicad_sch # what KiCad says is connected
netspec snap board.kicad_sch -o before.json # record connectivity, stably
netspec diff before.json board.kicad_sch # what actually changed
netspec check contract.py # adjudicate declared intent
netspec gate board.kicad_pcb # ERC/DRC at every severity
netspec guard board.kicad_sch -- <any tool> # snapshot, run, re-read, adjudicate$ netspec netlist tests/fixtures/good_ldo.kicad_sch
3 components, 3 nets
+3V3 C2.1, U1.2
GND C1.2, C2.2, U1.1
VIN C1.1, U1.3
Finds kicad-cli on PATH, via Flatpak, in a macOS bundle or a Windows install --
netspec doctor says which. A missing engine is an environment fault (exit 4), never a
finding about your design.
netspec never writes to a design file. It is the gate an editing tool's output has
to pass, not the tool.
A survey of ~90 KiCad AI/agent projects found three failure modes that ship today, and one property of KiCad's own defaults, all of which a netlist check catches and nothing else does:
- wires "added" that connect nothing, reported as success
- a pin-position helper off by a sign, silently wiring pin 2 where pin 1 was asked for — which reverses a polarised capacitor while passing ERC
pcb drc --schematic-parityat default severity reporting 0 problems on a board with 147, becausefootprint_symbol_mismatchdefaults towarning
Each is frozen as a regression test in tests/fixtures/, paired with
the netlist KiCad itself derived from it.
Two of those fixtures are the same circuit, wired correctly and wired backwards. KiCad's
ERC reports the same two violations for both -- reversing a polarised capacitor is
legal wiring, so no rule fires. netspec names it:
$ netspec diff polarized_cap_correct.kicad_sch reversed_polarized_cap.kicad_sch
NET CHANGES
~ Net-(C1-Pad2) +C1.2 -C1.1
FLOATING PINS
! C1.1 is no longer connected to anything
PIN SWAPS (a connection moved between pins of one part)
C1 on Net-(C1-Pad2): pin 1 -> pin 2 <- reverses a 2-pin part
WARNING: 1 pin swap(s) on a two-pin part. If any is polarised, it is now backwards,
and ERC will not tell you.
A contract is Python, not a data file -- more expressive, nothing to version, and no policy DSL to invent. Pins are named by function where the symbol offers one, because pin numbers are exactly what the known schematic-writer bugs corrupt.
from kicad_netspec import Spec, forbid, mirrors, net, polarity, through
board = Spec(
source="hardware/board.kicad_sch",
rules=[
net("VIN", ["J1.1", "C1.1", "U1.VI"]),
net("+3V3", ["U1.VO", "C2.1"]),
polarity("C1", plus="VIN", minus="GND"),
forbid("VIN", "GND"),
through("GND", "NT1", "GNDPWR"), # the only path between two grounds
mirrors("U2", "U3"), # two channels, wired the same
],
)netspec check imports and executes that module -- a contract is code. netspec diff
executes nothing.
guard does not care what did the editing:
$ netspec guard board.kicad_sch --contract contract.py -- claude -p "add a decoupling cap"
NET CHANGES
~ GND +C1.1 -C1.2
~ VIN +C1.2 -C1.1
PIN SWAPS (a connection moved between pins of one part)
C1 on VIN: pin 1 -> pin 2 <- reverses a 2-pin part
CONTRACT
FAIL C1 polarity: pin 1->VIN, pin 2->GND
pin 1 is on GND, expected VIN -- C1 IS REVERSED. ERC does not check this.
Exit codes are the contract: 0 clean, 1 a violation of the design, 4 an
environment fault -- a missing engine is never reported as a broken board.
A GitHub Action that comments on a pull request with the connectivity delta -- something no PCB team has and most would want, agent or no agent:
- uses: heibench/netspec@v0
with:
schematic: hardware/board.kicad_sch
contract: hardware/contract.py # optionalIt needs KiCad, so run the job in a KiCad container -- see
examples/workflows/connectivity.yml. The comment
is updated in place rather than stacked per push, and a missing engine reports that it
checked nothing rather than failing the board.
Installing the package registers a pytest plugin; no conftest needed:
from kicad_netspec.pytest_plugin import assert_net, assert_polarity
def test_the_bulk_cap_is_the_right_way_round(netlist):
assert_polarity(netlist("hardware/board.kicad_sch"), "C1", plus="VIN", minus="GND")Without KiCad on the machine, those tests skip rather than fail.
An MCP server ships in the mcp extra — six stateless verbs, one process per call:
pip install 'kicad-netspec[mcp]'{ "mcpServers": { "netspec": { "command": "netspec-mcp" } } }doctor · netlist · snapshot · diff · check · gate. The whole tool list costs
~834 tokens of schema, and CI fails if it exceeds 3,000 — surveyed KiCad MCP servers
range from 2,574 to 48,627, the largest spending a quarter of a 200K window before the
agent reads a file.
Every reply carries the exit code and what it means, so an agent cannot confuse "I could
not look" with "your board is broken". guard is deliberately not exposed: it runs
an arbitrary command, and an agent that can already run commands gains nothing from
being handed a shell by a verification tool.
- partspec — mechanical CAD parts vs declared intent
- netspec — PCB nets vs declared intent
- gerberdiff — fabrication output geometry
netspec does not do design review (see kicad-happy),
does not edit schematics or boards (see Konnect),
and does not place or route. It holds the ruler.
Apache-2.0