AI · electronics repair
AI applied to board repair, without the hallucinations.
Ask a general-purpose chatbot where the input fuse sits on a board it has never seen: it will answer anyway. WrenchBoard takes the opposite stance: an AI agent constrained by the real schematic, the real boardview and audited failure rules. It only cites what it can prove on your board.
See it on a real board01The shift
What AI actually changes in a repair shop
Day-to-day component-level repair isn’t limited by soldering skill: it’s limited by access to knowledge. The references exist: PDF schematics hundreds of pages long, boardviews, forum threads where the answer is in post 47, senior experience that retires with its owner. No search engine reasons over any of it.
That is precisely what a language model is good at: reading heterogeneous sources, cross-referencing them, reasoning over them in plain language. On one condition: that its answers are grounded in the board sitting on your bench, not in a statistical average of what it read during training.
02The architecture
The architecture: a constrained agent
WrenchBoard doesn’t bolt a chat box onto a viewer. The diagnostic agent is built into the tool, with 36 specialized tools it calls while reasoning:
- Reading the board: query the electrical graph compiled from the schematic, look up a component, list the failure rules matching the symptoms.
- Driving the boardview: highlight, zoom, dim, point at a pin, propose a measurement protocol the view follows step by step.
- Repair journal: log every measurement, compare before/after, record the confirmed cause at session end.
- Sensors: request a frame from the shop’s microscope when the reasoning needs to see the board.
Every reply passes through a sanitizer: a refdes the parser doesn’t know is flagged, not asserted. Hypotheses come out with their probability, sourced from the device’s audited rules, and when the agent reasons beyond the rules, on schematic topology alone, it says so.
03The memory
Layered memory, like a team
A senior tech’s worth is accumulated experience. The WrenchBoard agent mirrors that structure with four distinct memory stores: cross-device failure patterns and diagnostic playbooks (shared, read-only), the current device’s knowledge pack (read-only, built offline and audited), and the active repair’s notebook (read-write, re-read on every session resume).
The result: knowledge flows in the right direction. What a diagnosis confirms enriches the base; what the base contains is traceable and auditable; nothing rests on the model’s unpredictable memory.
04The depth
Choosing the reasoning depth
Not every fault deserves the same compute. At session open you pick the model tier: Haiku for quick diagnostics on known faults, Sonnet for everyday work, Opus (on the Pro plan) for the hard boards where long reasoning over the graph pays off. The same workflow, detailed on the motherboard diagnostics page, applies at every tier.
MNT Reform 2
05Shop AI
Shop AI, not demo AI
WrenchBoard’s bet: the AI that helps repair isn’t the one that impresses in a demo, but the one that’s wrong the least at the bench. That’s why the whole edifice, from the complete microsoldering software down to a single boardview highlight, is built around verifiability. Try a session on a covered device: the free plan is enough to judge for yourself.
Watch it work
The same agent, on a real board
This is not a mockup, it is the bench itself with its own interface, running on the MNT Reform pack exported from the engine: 487 real components and 2066 real pads. Every step below is what the engine answers on this board, from the device's rules for the symptom to the schematic graph behind the input fuse F1 and the board net it feeds. Here is what the agent does with a board that will not power up.
Eight tool calls, each one a real answer from the engine on this board, and not one component named from memory.
FAQ
Frequently asked questions
How is this different from asking ChatGPT or Claude directly?
A general-purpose chatbot reasons from memory, with no access to your board: it can invent a refdes or quote the schematic of a different model. The WrenchBoard agent is wired into the actually parsed board (boardview, the schematic's electrical graph, the device's failure rules), and every reference it cites is validated against that data before it's displayed.
Which AI model powers the agent?
Anthropic's Claude models, tier-selectable at session open (Haiku, Sonnet, and Opus on the Pro plan) depending on the reasoning depth you need. The project took 2nd place at Anthropic's "Built with Opus 4.7" hackathon.
What does "anti-hallucination" mean concretely?
Every agent reply goes through a sanitizer: cited refdes are checked against the components actually present in the parsed boardview (with typo tolerance), and proposed causes must come from the device's rules or the compiled electrical graph. An unverifiable reference is flagged as such, never presented as fact.
Does the AI take the measurements for me?
No, it replaces neither your probes nor your judgment. It structures the loop: probability-ranked hypothesis, step-by-step measurement protocol, interpretation of the values you report. It can also request a frame from your microscope to examine an area of the board.
Is my repair data used to train the model?
No. Sessions feed the platform's diagnostic memory (measurements, confirmed causes, playbooks), not language-model training. And the engine is source-available: a shop can self-host it and keep all of its data local.
Get started
Try Wrench Board free, in your browser.
Free plan, no card required. Open the app, pick a covered device, and start a diagnostic session with the agent.