Wrench Board: the AI workbench bringing Claude Opus to board-level repair

Wrench Board is an agent-native diagnostic workbench for board-level electronics repair, powered by Anthropic's Claude Opus 4.8. 'Board-level' means the work that happens after the obvious swaps — not replacing a screen or a battery, but finding and fixing the single failed component on the logic board itself: a dead power-management chip, a blown filter, a shorted capacitor.

Instead of a chat box you paste screenshots into, Wrench Board ingests two things about a device — its schematic and a 'boardview' file describing the physical layout — and builds a per-device knowledge pack. It then runs a Claude Opus diagnostic agent that doesn't just talk back: it pilots a 3D rendering of the real board, lighting up components, tracing power rails, and showing the technician where to put the probe, while the tech keeps the soldering iron in hand.

It is device-agnostic. Feed it a schematic and a boardview and it works the same on iPhone and MacBook logic boards, Android and Samsung phones, game-console motherboards, laptops, and single-board computers. Anything with a schematic and a boardview is fair game.

Its creator describes it as a senior microsoldering teammate: for an experienced technician, 'a second pair of eyes that never gets tired'; for an apprentice, 'a senior teammate who explains the boot sequence the tenth time, in their language, with their tools, without judgment.' It is a public, source-available project — free for personal study and for independent repair professionals to use as an internal tool when servicing their own clients.

Tens of millions of tonnes of electronics become e-waste every year, and a large share of it is recoverable at the board level. A phone or laptop that 'won't turn on' has very often suffered one tiny failure — a dead capacitor, a blown diode, a failed PMIC (power-management IC) — on a board that is otherwise perfect.

The catch is that finding and fixing that one part requires a microsoldering technician working under a microscope, and there are very few of them. Wrench Board calls this the 'last mile' of repair: the stage where a device is otherwise written off and thrown away.

And the hardest part of that last mile is not the soldering — it is the diagnosis. A modern logic board carries hundreds of components and dozens of power rails. Knowing which one failed, and why, is what separates a board-level repair from an expensive guess. That diagnostic bottleneck is exactly what Wrench Board sets out to attack.

Wrench Board was built by Alexis Chapellier, a working microsoldering technician operating as Repair Valley, an independent electronics-repair workshop. The origin story is refreshingly honest: for years he diagnosed boards by sending screenshots to Claude one at a time, manually, and pasting the answers into a paper notebook. He built the workbench he wished he had — one that could see the actual board, reason over the real schematic of that exact device, and remember what worked last time.

That work was recognised publicly. Wrench Board placed 2nd in Anthropic's 'Built with Opus 4.7' Claude Code hackathon, announced in June 2026, winning a prize of US$30,000 in Claude API credits. The award was reported by Anthropic's own announcement and corroborated by independent technology press.

One important clarification, in the interest of honesty: iTweak did not build Wrench Board and was not a hackathon entrant, and we are not affiliated with Anthropic or with Repair Valley. We are covering Wrench Board here as a notable public tool in our field, and we use AI-assisted board-level diagnostics as part of our own repair process. The award and the tool belong to its creator; the credit is his.

The workflow mirrors how a good board technician already thinks — it just gives each step an AI teammate.

First, it learns the device. You give it a device label (say, 'iPhone 11' or 'MacBook Air A2337') and, ideally, the schematic PDF. A four-stage 'knowledge factory' researches the device, distills a canonical list of its components and power rails, and writes a knowledge graph, a set of diagnostic rules, and a glossary — then a final quality-control pass audits the result. A focused pack lands in roughly two minutes; a full pack that also ingests a dense schematic can take 15 minutes or more, and it keeps building in the background while you work.

Then, it diagnoses with you. You open a live conversation, describe the symptom and your measurements in plain language, and the agent reasons out loud — but it also acts on the board. It highlights the suspect chip, isolates a power rail's entire cascade, draws arrows along the power chain from the connector to the CPU, dims everything irrelevant, and marks the exact pad to probe next. Every visual move is the AI showing its thinking on the actual board, not a stock diagram.

And it can see the board. The technician plugs a USB microscope or webcam into the bench; when the agent needs a closer look at a suspect component, it requests a still frame, reads the image, and folds what it sees back into its reasoning. The technician frames and focuses the shot — the AI asks for it on cue. Captures are saved with the repair, so an entire session can be replayed later, words and photographs together.

This is the single most important thing to understand about Wrench Board, and the reason it is trustworthy on a bench where a mistake is expensive. A general chatbot, asked about a board, will happily invent a component reference — confidently naming a 'U7' or 'C29' that may not exist. On a logic board, acting on an invented part means lifting the wrong chip and destroying a board that was repairable. Wrench Board is built so that cannot happen.

Its guarantee is defense in depth, in two layers. First, tool discipline: when the agent looks up a component, the lookup either returns a real, verified part or returns 'not found' with the closest real matches — it physically cannot fabricate one, and the agent is instructed to pick from the suggestions or ask the technician. Second, an outbound sanitizer: before any reply reaches the screen, the system scans it for component-shaped references and validates each against the actually-parsed board. Anything that doesn't resolve is visibly flagged rather than presented as fact.

The creator's phrase for this is 'precision over magic.' It is what lets a technician trust an AI's pointer enough to put a hot iron next to it.

What makes Wrench Board more than a clever chatbot is that it reasons over your exact board's electrical reality. When it ingests a schematic, Claude Opus reads it page by page and compiles an 'electrical graph' of the device: it sorts the nets into power rails versus logic, works out which components depend on which rails, and infers the boot sequence — the order in which rails come up at power-on. That is how the agent can say '+3.3V standby comes up first, then the main rail, then the CPU module' and trace a dead board backwards along that chain.

Underneath the AI sit two deterministic engines that never call a language model at all, which is precisely why their answers are auditable and repeatable. A forward simulator answers cause-to-effect questions: kill this component and the board blocks at this boot phase, with these rails dead. A reverse 'hypothesizer' answers effect-to-cause: given that the 3.3V rail is dead but a particular regulator is alive, here are the components that would explain it — ranked — and here is the single best measurement to take next to tell the top suspects apart.

That last part is the quiet genius of it: when several faults are equally plausible, the tool doesn't shrug — it tells the technician exactly what to probe next to narrow it down fastest. This is the difference between a generic 'phone repair' chatbot and a tool that genuinely understands the power topology of the board in front of you.

Wrench Board has a real, layered memory rather than a disposable chat history. Each device keeps its own knowledge, and each repair keeps the agent's own working notebook — its decisions, its measurements, its open questions. When a technician pauses a job and comes back, the agent re-reads what it actually wrote to re-orient itself, rather than trusting a fuzzy AI summary that could drift.

More importantly, it learns from confirmed outcomes. When a technician verifies a root cause, the agent records a durable 'field report' for that device — the at-fault part, the reported symptom, the failure mechanism. The next time anyone diagnoses the same model, those proven findings are recalled. In practice, the more real repairs a workshop does on a given device, the sharper Wrench Board becomes on that device.

To be precise about what that does and doesn't mean: this is recall of confirmed, field-proven findings plus a nightly self-improvement loop that tunes the deterministic engine against a fixed, human-curated benchmark and only keeps changes that measurably improve accuracy. It is not training or fine-tuning Claude on your data.

For a seasoned microsolderer, Wrench Board is a tireless second opinion that has the whole schematic in its head. For an apprentice, it is a patient senior who will explain the boot sequence for the tenth time without judgment. For the wider right-to-repair movement, it is a serious attempt to make board-level skill more shareable — built in the open, by someone who actually does the repairs.

Running it well takes two inputs per device: the device's schematic PDF and a boardview file describing the physical layout. With both, the agent can see the board, reason over its power topology, simulate faults, and narrow suspects to the next probe. With neither, it is still a knowledgeable assistant — but it is reasoning from general device knowledge, not from your board's actual causality. Sourcing those files (and the formats Wrench Board accepts) is a topic in its own right, which we cover in the rest of this series.

In the posts that follow we go deeper: how the knowledge factory and schematic ingestion actually work, how to set Wrench Board up on your own bench and run a first diagnosis, exactly what schematics and boardviews to feed it, and how the anti-hallucination design holds up under pressure.

Is Wrench Board free? It is source-available: free for personal evaluation and study, and free for independent repair professionals to use as an internal tool when servicing their own clients. Redistribution, hosted/SaaS deployment, and sublicensing require written permission from the author.

Is it open source? Not in the formal sense. The source code is public to read and study, but it is released under a proprietary, source-available licence — not an OSI-approved open-source licence. 'Source-available' is the accurate term.

What devices does it work on? Anything with a schematic and a boardview — iPhone and MacBook logic boards, Android and Samsung phones, game consoles, laptops, and single-board computers.

Do I need schematics and boardviews? For the advanced capabilities — visual board piloting, fault simulation, and rail-level reasoning — yes. It will still help as a general assistant without them, but it can't reason about your specific board's causality.

Does it replace the technician? No. It is explicitly a teammate, not an autopilot. The human keeps the iron, frames the microscope shots, takes the measurements, and makes the call; the AI accelerates the diagnosis and refuses to bluff.

Did it really win an Anthropic award? Yes — Wrench Board placed 2nd in Anthropic's 'Built with Opus 4.7' Claude Code hackathon, announced in June 2026, with a US$30,000 Claude API-credits prize. The tool and the award belong to its creator, Alexis Chapellier (Repair Valley); iTweak is not affiliated and did not enter.