Knowledge factory
Per-device knowledge packs, built offline. Scout researches the device, three Writers structure findings into a graph, rules, and glossary, and an Auditor gates the verdict — APPROVED, NEEDS_REVISION, or REJECTED.
Agent-native. Schematic-aware. Anti-hallucination by design. Built for board-level electronics repair.
The full demo we submitted to Anthropic for the “Built with Opus 4.7” hackathon — running on the MNT Reform motherboard (CERN-OHL-S, fully open hardware). Two minutes covering the diagnostic agent, the boardview canvas, the schematic compiler, and the live tool-driven workflow.
Independent repair technicians get blocked the moment a fault goes deeper than a screen swap. The references exist — but they live as PDFs nobody indexes, thirteen incompatible boardview formats, and forum threads from 2017 with “did you ever fix this?” at the bottom. No tool reasons over any of it the way a senior tech does.
Wrench Board changes that.
Each workflow is a self-contained surface — composable, scriptable, and exposed to the diagnostic loop as native tools the agent calls directly.
Per-device knowledge packs, built offline. Scout researches the device, three Writers structure findings into a graph, rules, and glossary, and an Auditor gates the verdict — APPROVED, NEEDS_REVISION, or REJECTED.
PDF schematics in, ElectricalGraph out. Page-by-page vision pass on Opus 4.7 — natively reasoned, not OCR-flattened — then merged, classified into power / logic / connector edges, and compiled with an inferred boot sequence.
Reads a knowledge pack and proposes simulator scenarios: cause → expected dead rails and components. Five validators gate the output — sanity, grounding, topology, pertinence, dedup — and a per-pack reliability score keeps the agent honest about its causal engine.
Live repair sessions on Anthropic Managed Agents, anti-hallucination on every reply. Four layered memory stores back the agent: cross-device failure patterns and reusable diagnostic playbooks (shared, read-only), the device's own knowledge pack (read-only), and a per-repair scribe notebook (read-write) it reads on resume. Tier-selectable at session open — Haiku, Sonnet, Opus.
Bring the file. The agent reads it.
Hover any pill to read what it does. Names lifted straight out of api/agent/manifest.py.
Drop your email to know when the public beta opens. No drip campaign, no growth-hack cadence — one email when the door is open.