AI on the repair bench, grounded in your documents.

Two minutes from documents to working knowledge

A senior technician who has seen a board a hundred times carries its architecture in their head. WrenchBoard builds that mental model explicitly, before the first probe touches the board.

The build is done by a small factory of specialized Claude roles: a Scout gathers what is known about the device, a Registry fixes the component inventory, three Writers work in parallel (one maps the board, one writes the clinical failure picture, one compiles the vocabulary), and an Auditor checks the result before anything is stored. From a schematic PDF and a boardview, the whole pass takes about two minutes.

The output is not prose, it is a queryable corpus on disk, per device. Every diagnostic session that follows reads from it and writes back to it: measurements, confirmed causes, open questions. Nothing restarts from zero.

The schematic becomes a graph you can query

A sixty-page schematic PDF is documentation for humans with hours to spend. WrenchBoard compiles it, page by page, using Claude vision, into an electrical graph: which nets exist, what class they belong to, which components sit on them, what the boot sequence looks like, with a quality report attached stating how much of the document was confidently read.

That compilation is what turns “search the PDF” into “ask the board a question”. What a specific component does, what shares its rail, what should be live at this boot phase: answers come from the graph in seconds, and they cite the graph, not the model’s memory.

Deterministic engines under the reasoning

The agent does not guess causal chains, it delegates them. Two pure algorithms sit at the core of the diagnostic stack and make zero language-model calls at runtime:

• The simulator advances the board’s boot sequence phase by phase and emits a timeline: which rails are dead, which components are unpowered, what blocks each phase.
• The hypothesizer takes a partial observation, the handful of measurements you actually made, and enumerates the one- and two-fault candidates that would explain it, ranked by how well each one fits.

Deterministic means checkable: the same measurements always produce the same candidates, and every causal chain can be verified by hand against the graph. The agent’s job is to reason on top of that scaffolding in plain language, at the level of the technician using it, not to improvise electronics.

A session at the bench, iron in hand

At the bench, the agent drives the boardview while you work: it highlights the pin to probe, traces the net under suspicion, annotates the measurement protocol on the board itself. If its reasoning needs eyes, it requests a frame from your USB microscope; you can also drop a close-up of a suspect chip into the chat.

You measure, it interprets, the loop continues. Every measurement is logged, every capture is persisted, and the whole session can be replayed afterwards: what was said, what was decided, what the agent actually saw. The full diagnostic workflow is structured the way a senior would run it, but the iron never leaves your hand.

The honest limits

WrenchBoard is at its strongest with a schematic and a boardview in hand; that is where verified inventories and deterministic chains operate at full capacity. Without those documents, it still builds a first knowledge pack from public sources, useful, but thinner. That gap is not a software limitation, it is a documentation-access problem, and it is the real subject behind “too complex to repair”. The canonical development target is the MNT Reform mainboard, chosen precisely because its documentation is open under the CERN-OHL-S license.