Twenty-one of them write to the design — place a part, draw a wire, route a track, pour a zone. The rest only read it back: the netlist, the board outline, a datasheet, the result of ERC or DRC, what a part costs. Asking it for a board is asking it to call these in sequence, and every call it makes is visible and undoable, the same as if you had clicked it yourself.
macOS, Windows and Linux · no account. Every platform.
Design a 5 V to 3.3 V regulator board — AMS1117-3.3, 10 µF in and out, screw terminals both ends.
I’ll find the regulator first, then the passives, and wire them before anything goes to the board.
Both capacitors go either side of it, then I’ll label the three nets.
What it can touch
01
There is no single tool that just does what you asked. Each one is a specific, narrow operation with its own arguments, so a request turns into a short, legible list of exactly what happened — not a paragraph explaining what it meant to do. The panel names each of them the way the list below does, in words; the identifier underneath is the model's business, not yours.
Ten of them, picked to show the range. There are 40 in all: 35 inside the schematic and board editor, and five more that look a part up, cost a bill of materials and check the board against what a factory can build. Fourteen of the 35 only read the design; the other 21 write to it — and three of those 21 only move the camera.
What it lands on
02
There is no shadow copy of the board that the agent edits and you review afterwards. When it places a part, a real footprint lands on this canvas with real pads — the same editor state either of you would be looking at.
See board layout →
Between the calls, the panel shows the working the model did to get there — the search it ran, what came back, the part it ruled out and why. It is not a summary written afterwards; it is what the model produced on the way to the calls, folded under a single line you can collapse once you have read it.
How a turn is put together →Measured, not claimed
03
There are 48 hand-labelled cases it is run against, and a case counts only because the engine, asked afterwards, agrees the board is in the labelled state — not because a tool call returned OK.
Two runs against the pinned deepseek-v4-flash: 35 of 48 on 7 August, 32 of 44 on 8 August — both about 73%. Placing parts is 8/8 and pouring copper is 4/4; repairing DRC violations is 2/6, and asking a clarifying question instead of guessing is 0/2. Those last two are printed for the same reason as the first two.
This is one page of a larger table, kept on its own page rather than folded into a claim here.
The whole table →macOS, Windows and Linux · no account. Every platform.