I'm not a programmer. I'm not a trained electrical engineer either — I'm self-taught, I think in block diagrams and signal flows, and I use tools for the math. When I need a filter at a specific frequency, I don't derive transfer functions by hand. I look it up, plug in the values, and verify it works.

That's exactly what LLMs can't do. They'll tell you to use a 10kΩ resistor and a 100nF cap for your lowpass filter, and it sounds confident, but nobody checked the math. There's no simulation. No verification. Just vibes. And if you're building something real, vibes don't cut it.

I kept thinking: the AI is actually good at the translation part. It understands what I want when I say "build me a bandpass filter centered at 440 Hz." It can pick a topology. It can even write a SPICE netlist. The part it can't do is verify that the netlist actually works. That's what simulators are for — and nobody had connected the two.

So I built SPICEBridge. It's an MCP server — basically a bridge between language models and ngspice. The AI handles the English-to-circuit translation, ngspice handles the physics, and a structured tool layer handles everything in between. 28 tools covering simulation, measurement, design equations, tolerance analysis, schematic generation, and KiCad export.

The key design decision: don't trust the AI with math. The design equation solver computes component values deterministically from your specifications and snaps them to standard E-series values. The AI orchestrates the workflow — it doesn't calculate resistor values. Same philosophy as choosing a 555 timer over an Arduino for a simple timing circuit. Use the right tool for each job.

I built the initial system in under 24 hours using the same multi-AI workflow I used for CodeDocent: one Claude instance for thinking and planning, other AI models for adversarial code review, and Claude Code for implementation. Then I spent another two days on a full security audit — 8 automated scanners plus reviews by ChatGPT, Gemini, DeepSeek, Grok, and Kimi. 56 findings across 8 fix batches. 771 automated tests.

It's GPL v3, pip installable, and the hosted server is live if you want to try it from Claude.ai without installing anything. Point it at a circuit problem and let the AI actually verify its own work for once.