Microsoldering software that reasons before it reworks.

What software should bring to a microsoldering bench

Screen and battery swaps don’t need software. It’s at component level (shorted power rails, missing voltages, failing BGAs, corrosion after liquid damage) that tooling makes the difference between a one-hour repair and a board that sits on the shelf for three weeks.

Good microsoldering software has four concrete jobs:

• Read the board’s documentation: boardviews in mutually incompatible formats, PDF schematics hundreds of pages long, and the cross-references between the two.
• Locate physically: go from a refdes (U7, R49, F1) to its real position on the top or bottom side, with the nets that reach it.
• Structure the diagnosis: turn a symptom (“board dead, no LED”) into an ordered sequence of measurements instead of probing at random.
• Accumulate: what you found on one board should serve the next one.

WrenchBoard is built around those four jobs, with one particularity: an AI agent is wired into every one of them.

Reading the board: boardview and schematics in one tool

WrenchBoard ships a boardview viewer that opens thirteen formats natively, from KiCad .kicad_pcb to the historical .brd, plus .fz, .cad, .bdv, .asc, .tvw and the XZZ-encrypted .pcb. The board renders in an interactive canvas: refdes search, net highlighting, top/bottom flip, component-type filtering, on desktop and touch alike. One format earns a special mention: the XZZ .pcb export carries the factory’s own test values, the expected resistance and voltage on each net, so the agent can compare your reading against what the board is supposed to show.

On the schematic side, the ingestion pipeline turns a PDF into an electrical graph: every page goes through a vision model (Claude), components and connections are extracted, merged, classified into power, logic and connector edges, then compiled with an inferred boot sequence. The output is not an annotated image. It’s a queryable data structure you can run “what happens if F1 is open?” against.

From symptom to measurement protocol

The heart of the workflow: you describe the symptom in plain language, the diagnostic agent crosses the device’s known failure rules with the electrical graph, and proposes a probability-ranked hypothesis, then a step-by-step measurement protocol: continuity across the input fuse, voltage at this test point, diode-mode reading on that rail.

Every measurement you report lands in the repair journal and tightens the hypothesis. The agent doesn’t describe the board. It drives it. When it asks you to probe R49, the boardview zooms in, highlights the part and dims everything else.

For structuring fault-finding on power trees specifically, see the motherboard diagnostics page, which details the fault-tree method the agent applies.

Anti-hallucination: the constraint that changes everything

A language model left free with its answers will eventually invent a capacitor. WrenchBoard enforces the opposite: every component reference the agent cites is validated against the actually parsed board. An unknown refdes is flagged, never asserted. Failure rules come from a per-device knowledge base, built offline and audited, not from the model’s memory.

That’s the difference between asking a general-purpose chatbot for advice and working with a trade tool: the first will always answer something; the second only cites what it can prove on your board.

Accumulating: memory across repairs

Every session leaves a structured trace: measurements, confirmed cause, narrative summary. Recurring failure patterns and diagnostic playbooks are shared across devices read-only; the active repair’s notebook is read-write. When a similar board comes back, the agent re-reads what already worked, like the senior colleague you’d ask “have you seen this fault before?”.

Donor stock, microscope, and a touch-driven bench

Three more surfaces the agent reaches into. Donor stock: declare your dead boards as donors and WrenchBoard indexes their harvestable parts, so when a repair needs a replacement the agent searches your own pile by electrical signature (type, value, package, MPN) before you order anything. Microscope capture: it can pull a still frame from your USB scope or webcam mid-session to look at what you’re seeing. And the boardview is fully touch-driven on phone and tablet, with pinch-zoom, pan and tap-select, so the diagnostic screen works on the same device you keep next to the hot-air station.

In the browser, or on your own metal

WrenchBoard runs in the browser with no install and a free plan to get started on covered devices. The diagnostic engine is source-available: a shop that wants to keep schematics and history local can self-host it for its own bench. The project took 2nd place at Anthropic’s “Built with Opus 4.7” hackathon in April 2026, and the code is on GitHub.