The Problem
Turning circuit requirements into a physical board layout that works in the real world is a hard spatial reasoning problem. Components must be placed and connected in 2D and 3D space while respecting electrical behavior, heat dissipation, mechanical enclosures, and manufacturing limits. This is reasoning across multiple domains at once — electronics, geometry, physics, and manufacturing constraints — all interacting in a single design. Much of this knowledge isn’t centralized; it’s scattered across datasheets, reference designs, application notes, and engineer intuition built over years. At its core, PCB design is a multi-objective optimization problem. The system must decide where parts go and how traces are routed, balancing short connections, safe spacing, heat flow, and signal quality — all while avoiding collisions and dead ends. Small changes in layout can break the design or make it impossible to manufacture.Why Existing AI Approaches Don’t Work
Most attempts to apply AI to PCB design fall into four categories, and each has a fundamental limitation:Code-First Tools
Some tools require engineers to describe circuits in a programming language. The problem: electrical engineers don’t want to write code to design circuits. Adding a code layer between the engineer and the design increases cognitive load without solving the underlying spatial reasoning problem. Engineers think in schematics and layouts, not in function calls.Browser-Based Tools
Cloud-hosted design tools require uploading proprietary designs to someone else’s server. No serious engineering team is going to put their IP on a third-party platform they don’t control. This is a non-starter for any proprietary or production workflow.Routing-Only Tools
Some tools focus exclusively on autorouting — taking a placed board and finding trace paths. Routing is one step in a much larger pipeline. It doesn’t solve component selection, schematic correctness, placement optimization, design rule validation, or iteration speed. Solving routing alone doesn’t make the overall design process meaningfully faster.GUI Automation
Some approaches try to make AI control existing EDA tools by scripting mouse clicks and keyboard shortcuts — essentially screen-scraping a design tool. This is fragile, slow, and fundamentally limited to whatever the GUI exposes. The AI can’t reason about the design at a structural level; it can only mimic a user clicking buttons. When the host tool updates its interface, everything breaks.The Common Failure Mode
These approaches fail because PCB design isn’t a single-step problem. Code-first tools add cognitive load, cloud tools block serious teams, routing-only tools don’t address correctness or iteration speed, and GUI automation is limited by the host tool’s interface. Treating PCB design as one narrow task instead of a multi-domain reasoning problem leads to tools that are impressive demos but don’t hold up in real workflows.How Trace Is Different
We didn’t try to teach AI how to click through existing PCB tools or bolt a chatbot onto a legacy interface. Instead, we built a new foundation.A Custom Circuit Representation
Trace uses a purpose-built text syntax for circuit files (trace_sch for schematics, trace_pcb for layouts). Language models are exceptionally good at reading and editing structured text. By creating a format designed for AI to work with natively, the model can make precise, targeted edits to any part of a design — adding a component, adjusting a connection, modifying a footprint — without needing to interpret a GUI or reverse-engineer a binary format.
This is the key architectural decision. Other tools try to make AI interact with circuit designs through screen coordinates, GUI automation scripts, or programmatic APIs that weren’t designed for language models. Trace inverts this: the file format itself is the AI interface.
The AI Edits Text, You Edit Visually
The AI works on the text representation while you work in the graphical editor. Both agent and user manipulate the same design through their natural interface, without breaking or redesigning the existing workflow. There’s no mode switch, no export/import cycle, no “AI view” vs “human view.” One design, two interfaces, real-time sync. This matters because it means the AI doesn’t need to “see” your screen or figure out where to click. It reads and writes structured text — the thing language models are best at — while you keep your normal visual workflow.Multi-Domain Reasoning, Not Single-Step Automation
Trace combines high-level reasoning — choosing components, planning circuit topology, understanding design intent from natural language — with specialized solvers that handle physical layout optimization. The system alternates between these layers:- Intent understanding — Parse what the engineer wants from natural language, datasheets, or reference designs
- Component selection — Choose parts based on availability, cost, specs, and compatibility
- Schematic generation — Build correct circuit topology with proper connections
- Placement optimization — Position components considering signal flow, thermal requirements, and mechanical constraints
- Routing — Route traces with impedance control, length matching, and DRC compliance
- Validation — Run electrical and design rule checks, simulate behavior, catch errors before manufacturing
Multimodal Input
Engineers don’t start from a blank page. They start from datasheets, reference designs, application notes, hand-drawn sketches, and existing boards. Trace accepts all of these as input — attach a datasheet PDF, a photo of a hand-drawn schematic, or a screenshot of a reference design, and the AI extracts the relevant information and incorporates it into your design. This closes the gap between “I found a reference circuit in a datasheet” and “that circuit is now in my schematic with real, in-stock parts.”Real-Time Supply Chain Awareness
The AI doesn’t suggest theoretical parts. It queries live distributor inventory across JLCPCB, LCSC, DigiKey, Mouser, Texas Instruments, and Amazon during the design process. If a part goes out of stock or end-of-life, Trace suggests pin-compatible alternatives. Component selection is grounded in what you can actually buy and assemble today.Spatial Reasoning at Scale
As spatial and multimodal models improve, this kind of cross-domain reasoning is becoming tractable in software for the first time at useful accuracy. Trace pairs language model reasoning with graph-based algorithms that handle spatial placement and routing — combining the flexibility of natural language interaction with the precision of constraint solvers. The knowledge required for good PCB design — thermal behavior, EMI considerations, manufacturing tolerances, signal integrity rules — isn’t centralized anywhere. It’s scattered across thousands of datasheets, application notes, and years of engineering intuition. AI is the first tool capable of synthesizing all of that context simultaneously during the design process.Built on KiCad 10, No Lock-In
Trace is built on KiCad 10, the most widely used open-source EDA suite. Your designs are stored in standard KiCad formats (.kicad_sch, .kicad_pcb). You can open any Trace project in vanilla KiCad, and open any existing KiCad project in Trace.
There’s no proprietary format you can’t escape. If you stop using Trace tomorrow, your designs are still yours in a format the entire industry understands. The AI intermediate format (trace_sch / trace_pcb) is a compile step — it converts to standard KiCad files automatically, and the KiCad files are what you keep.
Desktop-First, Privacy-First
Trace is a native desktop application. Your designs live on your machine. The AI assists your workflow through a streaming connection, but your files never leave your control. There’s no cloud workspace, no shared storage, no “someone else’s server.” Enterprise teams can deploy private backend instances for complete isolation.Trust and Control
AI making changes to a circuit design is high-stakes. A wrong connection or a missing decoupling capacitor can mean a dead board and weeks of delay. Trace is designed around the assumption that engineers need to trust and verify every change.Plan Mode
For complex tasks, the AI creates a detailed step-by-step plan and presents it for your review before executing anything. You can approve, adjust, or reject. This means you understand what’s about to happen before it happens.Version History
Every AI edit is versioned. You can roll back to any previous state — before a specific change, before an entire conversation, or to any point in the design’s history.- Local history — Versions are stored in a
.history/directory within your project using git. This works entirely offline and persists across sessions. Every AI edit automatically creates a commit, so you have a complete timeline of changes - Cloud sync — Versions are also saved to Trace’s cloud for cross-device access and backup
- Per-message rollback — Each assistant message in the chat has an undo button. Clicking it restores the design to the state before that message’s edits were applied
- Non-destructive — Rolling back creates a new version, so you can always go forward again. No history is ever deleted