The harness for
electronics engineering.
A browser-based PCB workspace where AI agents place, wire, route, and check — directly in your native ECAD files (KiCad, Altium, Cadence, Siemens) — and keep working after you close the tab. You express intent in hardware terms. The harness runs the AI underneath.
Every tool contributes. One loop ties them together.
A product isn't just a board — it's requirements, parts, simulation, mechanical, sourcing, and whether the thing can actually be built. A finished PCB is thousands of trade-offs across all of them. The harness is the connective tissue: one loop agents can act on, constraints checked where they live, traceability through Git — so rapid iteration is safe and nothing is lost between tools.
Circuitly
browser-based harness · regulated-ready
LLMs ⇆ electronics
agents in the loop
Getting value from LLMs is its own discipline.
Context, experiments, model choice, orchestration, guardrails — and none of it is electronics engineering. The harness absorbs that burden, so your team designs hardware instead of operating AI.
Engineering-grade precision
Electronics doesn't forgive "almost right." Every agent edit runs against your design rules, gets checked before merge, and lands as a reviewable diff. Precision is enforced by the harness — not hoped for from the model.
Model ops, out of the box
Which LLM reasons best over datasheets? When does a deterministic solver beat a model at routing? The harness runs that experimentation continuously — benchmarking LLMs and deterministic tools per electronics use case, re-tuning as they evolve. You never touch a model picker.
Un-siloed by design
Schematic, layout, sourcing, supply risk, PLM, fab — each lives in its own tool, and context dies at every handoff. The harness is the connective layer: one workspace where context carries across the whole chain.
Four words the market uses loosely. We use them precisely.
Agent, sub-agent, orchestrator, harness. They're layered — each builds on the one before. Defining them clearly is part of the pitch: it makes our claim legible.
An LLM that owns its own loop — it picks a tool, sees the result, and adjusts, until the job is done. Not a fixed script.
A focused agent a parent spins up to handle one piece of the job. It runs its own loop, then reports back — kept simple and bounded.
The coordinator that decides which agents run, in what order, and combines their work into one result.
Everything around the agents that makes them usable and safe: the workspace they act in, tools, memory, guardrails, review, rollback. The layer you actually work at. This is Circuitly.
Prompt, plan, build, merge.
You express intent in hardware terms. The harness runs the AI underneath.
Prompt
Say it in hardware terms — "add gigabit Ethernet: PHY, magnetics, and length-matched RGMII." No prompt engineering, no context wrangling. That's the harness's job.
Plan
Agents propose before they touch anything. Datasheets, supply data, library rules, and your existing files come in as context — you approve the plan.
Build
Agents place, wire, route, and check in your real files — in parallel, server-side. Close the tab; they keep working while you sleep.
Merge
Come back to finished branches. Review visual diffs, roll back anything, merge what's right. Git underneath — you in the loop.
One prompt. A tree of agents.
A single agent handles a single task. Real hardware work needs many, coordinated. In Circuitly, an agent spawns another that runs its own loop, does its piece, and reports back — nesting as deep as the job requires. The parent waits; the child works; the result flows back up.
Most "AI for X" tools route inside one loop. Coordinating a tree of independent agents as one governed system is what turns a high-level request into real, reviewable work.
We're precise about what counts: handing routing to DeepPCB is an external job the harness waits on — not an agent.
Real files. Real tools. No lock-in.
An agent is only as useful as the place it can act. So we built the substrate: a real PCB design tool, in the browser, git-native — where agents read and write the actual files.
Native ECAD files
Agents read and write real KiCad, Altium, Cadence, and Siemens files. No proprietary intermediate format.
Git-native
Every agent edit is a commit on a branch. Full traceability, visual diffs, one-click rollback across the whole tree.
Guardrails built in
Plan-then-edit review, bounded sub-agents, and design-rule checks on every change. Nothing merges without you.
Self-hosted for enterprise
Deploy the harness inside your own infrastructure. Your designs never leave your network.
Not another tool. The layer above them.
Getting value from LLMs is its own discipline — context, experiments, orchestration, guardrails — and none of it is electronics engineering. The harness absorbs it.
"AI that helps you draw PCBs"
The workspace agents work in
Frequently asked questions
What is Circuitly?
How is Circuitly different from using ChatGPT or an AI copilot for circuit design?
Which ECAD tools does Circuitly support?
What can the agents actually do?
Do the agents really keep working when I close the tab?
Which AI models does Circuitly use?
Can we self-host Circuitly? Is our design data safe?
Put your agents in a harness.
See agents work in your files on a 30-minute call, or join the waitlist for early access.