Essay · 2026-09-06 · Stefan Jančić — ALT TAB
ALT TAB Circuit Copilot Demo KiCad AI Plugin vs GPT-6 Astra: Full PCB Design From One Prompt
GPT-6 Astra does PCB layout by driving KiCad's interface from outside. We build inside the editor. Both work — here's the difference, and why it matters for your board.
When OpenAI announced GPT-6 Astra, one of the capabilities they chose to lead with was PCB layout in KiCad — placing components, routing copper, turning a schematic into manufacturing files.
“Integral to every electronic device today, PCB layout is a manual task and common source of latency in the electronics design process. Accelerating it means freeing engineers to invent, optimize, and test their next idea at a significantly higher cadence.”
OpenAI — Introducing GPT-6 Astra
We build an AI copilot that lives inside KiCad. People started emailing us within the hour. So let’s talk about it properly.
First thing to say: this is good news, and we’re not going to pretend otherwise.
The largest AI lab in the world just told every investor, journalist and hardware engineer that AI-assisted PCB layout is real and worth doing. We have spent a long time explaining that to people one at a time. That argument is now settled, by someone with a much bigger megaphone.
The interesting question isn’t whether AI should touch your board. It’s how it touches it.
What Astra actually does
Read the announcement carefully and the architecture is right there. Alongside Astra, OpenAI shipped an updated Codex harness to “significantly improve the speed of computer use” — 1.9× faster than GPT-5.6 Sol.
Computer use. That’s the mechanism. Astra operates KiCad the way a person does: it looks at the screen, moves the pointer, clicks, types, looks again. From outside the application, through the interface. The demo video — a 15-second condensed playback of a much longer run — shows exactly that, and it is genuinely impressive to watch.
That approach has a real superpower: it works on anything. No API needed, no plugin, no integration. The same agent that drives KiCad can drive Altium, or your ERP, or a fab portal. For a general-purpose model, that generality is the whole point, and it is the right call for them.
It also has costs, and if you design boards for a living they’re worth knowing.
What driving from outside costs you
None of this is a knock on the model. These are properties of the interface, not the intelligence behind it.
Screenshots are a lossy channel
Your board is a precise data structure — netlists, footprint geometry, layer stackup, clearance rules. Rendered to pixels, most of that is gone. The agent sees a picture of your board, not your board.
A click has no transaction
When a dialog opens unexpectedly or a click lands two pixels off, there is no clean rollback. Public FPV-controller runs hit duplicate windows, kill -TERM, and stale DRC caches.
Every run starts from zero
One public F405 six-layer FPV run reported ~52 million tokens across 3 h 13 m 56 s — much of it rediscovering KiCad, writing a placer, a router, a DRC checker. Session ends. Next board, same bill.
Your design leaves the building
Screenshots and file contents go to a remote model. Fine for plenty of work. Under NDA, that’s a conversation with legal.
“Manufacturable” is doing a lot of work. A board that passes DRC is one a fab will accept. Whether it functions — switching node, ground return, crystal placement — is a different question. The FPV author called the result a research prototype, not yet sent out. That’s the honest state of the art.
What we do instead
ALT TAB Circuit Copilot runs inside KiCad’s own process. Not a terminal beside it. Not a screenshot loop. A panel in the editor, holding the same board object the editor holds.
It edits the board, not a picture of the board
Around sixty tools operate the real data structure: place symbol, route track, add via, assign footprints, set the layer stack, run ERC, run DRC, export Gerbers, export STEP. Each one executes in milliseconds, deterministically, and costs zero tokens.
You watch it work, and you can stop it
Every edit is an atomic write to the real file. Diff panel, checkpoints, accept and reject per change, Continue if the connection drops. No black box that hands you a ZIP three hours later.
Vision is one input, not the only channel
We render schematic and board for a vision model so the agent catches stranded parts and silkscreen collisions — on purpose, not as the entire interface.
It learned from boards that got built
33,502 open-source KiCad projects: decoupling caps median 9.5 mm from IC, crystals 10.5 mm, 64% of connectors within 6 mm of the board edge, 98% of placements orthogonal. Measurements, not opinions.
Your design can stay on your machine
Run a local model and nothing leaves your computer. Not the schematic, not the board, not a screenshot.
We’re not racing GPT-6. We’re the thing it should plug into.
We are not trying to out-reason a frontier model. That would be a foolish fight. What we build is the harness — tools, safety rails, file handling, visual feedback, recovery. Bring whichever brain you like. Circuit Copilot already speaks to OpenAI, Anthropic, Google and local models through the same tool layer.
Astra spent hours rediscovering how to talk to KiCad. In our plugin that part is already done — for every model, on every run, in milliseconds. The model is the brain. We’re the hands, and the hands already know where everything is.
What’s real today, and what isn’t
We’d rather you trust us in six months than be impressed today.
Shipping now
- Native plugin in schematic and PCB editors
- ~60 design tools, atomic writes, checkpoints, diff, Continue
- ERC and DRC with automatic validation rounds
- Visual inspection of rendered schematic and board
- Hierarchical sheets, up to 12 copper layers, STEP export
- OpenAI, Anthropic, Google, or a local model
- Placement conventions from 33,502 real projects
In progress
- Ohm 2.0 — KiCad-specialised, local hardware
- Net-aware placement from mined rules
- Independent DRC as a second opinion
- Longer autonomous runs for a full board in one pass
We will not claim this replaces you. Every schematic, footprint, netlist and board needs an engineer’s review before fab. The Astra results need one too — their authors said so.
Watch a real run, uncut
Compare it to a 15-second condensed playback. Different approach, different trade-offs. We think the difference is worth your attention.
Sources and references
- OpenAI — “Introducing GPT-6 Astra” openai.com/index/gpt-6-astra
- @gogofly23 — F405 six-layer FPV flight controller run (~52 M tokens, 3 h 13 m 56 s) x.com/gogofly23
- KiCad EDA — GPL-3.0-or-later kicad.org
Short quotations from the OpenAI announcement are used for commentary and review, with attribution and links to the original. We have not reproduced OpenAI’s video or images. All trademarks belong to their respective owners. ALT TAB Circuit Copilot is an independent project by STEFAN JANČIĆ PR RAČUNARSKO PROGRAMIRANJE ALT TAB VRANJE. It is not affiliated with, sponsored by, or endorsed by OpenAI, Alibaba Cloud, or the KiCad project. If KiCad saves you money, consider supporting it.