Skip to the content
6 layers276 × 93.5 mmMnt reform motherboard, the printed circuit board in 3D

Mnt reform motherboard

Junkz3/wrench-board259

A senior microsoldering technician, available to every repair shop from the seasoned pro to the apprentice. Powered by Claude Opus 5

Size
276 × 93.5 mm
Copper layers
6
Thickness
1.6 mm
Parts
505
Nets
647
Pads
2,109
Tracks, vias
10,953, 1,081
Pours
32
Routed
Yes
KiCad
KiCad 7 (file 20221018)
Created
21 Apr 2026
Updated
20 Jul 2026

Overview

Mnt reform motherboard is an open-source printed circuit board by Junkz3, from the Junkz3/wrench-board repository on GitHub. It is a 6-layer board of 276 × 93.5 mm with 505 parts on 647 nets, routed with 10,953 tracks and 1,081 vias.

It is built around the TLV62568DBV, AP22815AWT and SP3012-06UTG, with the PMEG60T10ELRX, PMV50EPEAR and SI7850DP among its other parts. Its parts include 178 resistors, 164 capacitors, 19 diodes, 14 transistors and 10 inductors. It belongs with the computers and single-board computers boards of this library.

Main components

12

README

README.md

Wrench Board mascot

English · Français · 中文 · हिन्दी

Wrench Board

GitHub stars Python 3.11+ License: source-available Powered by Claude Opus 4.8

Agent-native diagnostic workbench for board-level electronics repair, powered by Claude Opus 4.8. Right-to-repair, built in the open, by the people who actually do the repairs.

🥈 2nd place at Anthropic's Build with Opus 4.7 hackathon, April 2026.

📺 Demo video (3 min): https://youtu.be/OZ2D_p82z6w

Try it online: open a boardview in the browser, free. No account, no install: .brd, .fz, .tvw, .pcb and more, parsed natively.

Wrench Board: boardview + diagnostic agent on an MNT Reform motherboard

What it is

Tens of millions of tonnes of electronics end up as e-waste every year. A large share of that is recoverable at the board level: a dead capacitor, a blown diode, a bad PMIC, but only a microsoldering technician can find and fix it. We are the last mile of repair before the landfill, and there are not many of us.

Wrench Board is a senior microsoldering teammate built for that last mile. For the seasoned tech, it's a second pair of eyes that never gets tired. For the apprentice, it's a senior teammate who explains the boot sequence the tenth time, in their language, with their tools, without judgment. It ingests a schematic PDF and a boardview, builds a per-device knowledge pack in two minutes, and runs an Opus 4.8 diagnostic agent that pilots the board visually (highlights pins, traces nets, simulates failures) while the technician keeps the iron in their 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.

The bet is precision over magic. The agent is not allowed to invent a reference designator. Every refdes it utters originates from a tool lookup, and a server-side sanitizer wraps any token it cannot verify before the text reaches the screen. The deterministic engines underneath produce verifiable causal chains, not vibes.

Why it exists

I've been a microsoldering technician for three years. For most of that time, I sent screenshots to Claude one at a time, manually, and pasted the answer into a paper notebook. I built the workbench I needed.

How it works

Four orthogonal workflows feed a single on-disk corpus per device under memory/{slug}/:

  • Knowledge Factory: four Claude personas (Scout, Registry, Writers, Auditor) build a verified repair pack from a device label in ~2 minutes. The three Writers (Cartographe / Clinicien / Lexicographe) run in parallel and share a cache-warmed prefix to amortize the long shared input across writers.
  • Schematic Ingestion: Opus 4.8 vision compiles a PDF schematic, page by page, into a queryable ElectricalGraph: nets classified, boot sequence inferred, quality report attached.
  • Diagnostic Agent: an Anthropic Managed Agent per device, with a four-store layered memory (global-patterns, global-playbooks, device-{slug}, repair-{repair_id}), pilots the boardview through 17 bv_* tools and queries the pack, schematic graph, measurements, validations and technician profile through 27 more: 44 custom tools declared in api/agent/manifest.py. The agent never fabricates a refdes: tool discipline plus a post-hoc sanitizer.
  • microsolder-evolve: four overnight search loops, one per surface: the deterministic simulator + hypothesize engines (sim), the schematic compiler (pipeline), the schematic vision pass (pipeline-vision), and the diagnostic agent itself (agent). Each loop proposes patches against an oracle benchmark and either keeps them (evolve:-prefixed commit) or reverts. The loops have been running and shipping improvements while I work on other things.

Wrench Board: repair dashboard with knowledge artefacts and diagnostic threads

Files + Vision: the agent can ask to see

A microsoldering diagnosis lives or dies on what the probe is touching right now, and a chat box can't carry that. The technician plugs a USB microscope or webcam into the workbench and the agent requests a frame on demand through the cam_capture tool, reads the image, and feeds it back into its reasoning. The technician can also drop a macro shot or a close-up of a suspect chip into the chat at any time. Captures and uploads are persisted under the repair so a session can be replayed end-to-end: words, decisions, and the actual photographs the agent looked at.

This closes the loop the screenshot-pasting workflow never could: the agent stops guessing what the board looks like and starts seeing it, on the technician's cue, on the technician's optics.

Under the hood

  • Backend: Python 3.11+ / FastAPI / native WebSocket / Pydantic v2 / pdfplumber. No build step, no bundler.
  • Frontend: vanilla HTML + CSS + JS, OKLCH design tokens, D3 v7 for the knowledge graph and Three.js r128 (WebGL) for the boardview. Inline SVG icons. No framework.
  • Models: Claude Opus 4.8 (heavy pipeline writers, schematic vision, deep diagnostic tier), Claude Sonnet 4.6 (Scout, Registry, Mapper, Lexicographe, normal tier), Claude Haiku 4.5 (intent classifier, phase narrator, coverage gate, fast tier).
  • Memory: per-device Anthropic Managed Agents memory stores. The agent self-orients across sessions by reading its own scribe notebook (state.md, decisions/, measurements/, open_questions.md) instead of relying on an LLM-generated resume.
  • Boardview: 16 parsers in api/board/parser/, dispatched by extension: KiCad .kicad_pcb, OpenBoardView Test_Link .brd, KiCad-boardview BRD2, plus .asc .bdv .bv .bvr .cad .cst .f2b .fz .gr .pcb .tvw. Adding a format = one new file.
  • Tests: 2 600+ fast tests (~1 min) plus a @slow accuracy-gate suite, including 10 deterministic invariants on the simulator + hypothesize engines and frozen-oracle gates.
  • Tooling: make doctor runs 8 local health checks (env, packs, parsers, camera) for atelier deployment. make eval-all orchestrates the four eval surfaces (simulator, pipeline, vision, agent) with cross-skill regression detection. make tools-inventory writes a local agent-manifest index for offline review.
  • Anti-hallucination: defense in depth, two layers. (1) Tools return {found: false, closest_matches: [...]} for unknown refdes; the system prompt instructs the agent to pick from suggestions or ask the user. (2) api/agent/sanitize.py scans every outbound text for refdes-shaped tokens (\b[A-Z]{1,3}\d{1,4}\b) and wraps any unverified match as ⟨?U999⟩ before it reaches the technician.

Two pure-sync deterministic engines (simulator.py, hypothesize.py) sit at the core of the diagnostic stack. The simulator advances phase-by-phase over a boot sequence and emits a timeline of dead rails, dead components, and the cause of blocking per phase. The hypothesizer takes a partial observation and enumerates 1- and 2-fault refdes-kill candidates that explain it, ranked by F1 against the observation. Neither calls an LLM at runtime.

The diagnostic agent has two interchangeable runtimes: managed via Anthropic Managed Agents, direct via the Messages API. Managed is the default and the production path; direct serves as a fallback when the MA beta is unavailable and as an on-disk inspection harness during development. The WebSocket protocol is identical so the frontend doesn't know which one is running.

Roadmap: Community Evolution Loop

Wrench Board runs locally. Each technician's instance can improve its deterministic simulator against their own field cases. When the evolve loop discovers a rule that holds up, it surfaces a candidate pull request to the upstream repo. Right-to-repair, built in the open, by the people who actually do the repairs.

Quickstart

bash
git clone https://github.com/Junkz3/wrench-board
cd wrench-board
make install          # create .venv and install deps (incl. [dev])
cp .env.example .env  # then fill in ANTHROPIC_API_KEY
make run              # uvicorn --reload on http://localhost:8000

On the first make run in Managed Agents mode (default), the start script prints a one-screen warning describing what it is about to create on your Anthropic account (1 environment + 3 tier-scoped agents, idle, no cost until used) and waits 5 seconds for Ctrl+C before bootstrapping. The IDs land in managed_ids.json (gitignored) and subsequent runs go straight to uvicorn.

Fallback to direct mode if the Managed Agents beta is unavailable on your account, no bootstrap, plain messages.create tool loop:

bash
make demo-fallback
# or: DIAGNOSTIC_MODE=direct make run

License & credits

Source-available under a proprietary license. See LICENSE. Free for personal evaluation, study, and local use. Independent electronics repair professionals may also use it as an internal tool when servicing their own clients (commercial remuneration OK), with no separate licence needed. Redistribution, hosted SaaS deployment, sublicensing, and any use for training competing AI / ML models still require written permission (contact: alexis@repairmind.co.uk). Dependencies are MIT / Apache 2.0 / BSD only. The MNT Reform motherboard used as the canonical test target is CERN-OHL-S-2.0. Built solo at Repair Valley, an independent electronics repair workshop.

Contributing

Wrench Board is open to contributors who care about right-to-repair. Field reports, new boardview parsers, simulator rules: open an issue or a PR.

Parts

505
QtyValueFootprintReferences
730.1uFCapacitor_SMD:C_0603_1608MetricC11, C12, C13, C101, C103, C105, C106, C107, C108, C109, C110, C113, C115, C118, C119, C135 +57
400Resistor_SMD:R_0603_1608MetricR1, R101, R104, R105, R106, R107, R108, R109, R111, R116, R124, R125, R126, R129, R130, R137 +24
2310kResistor_SMD:R_0603_1608MetricR2, R45, R54, R113, R115, R117, R118, R119, R128, R161, R168, R169, R170, R171, R172, R196 +7
2310uFCapacitor_SMD:C_0603_1608MetricC34, C37, C46, C47, C63, C112, C120, C131, C132, C133, C134, C137, C150, C155, C164, C167 +7
171uFCapacitor_SMD:C_0603_1608MetricC1, C4, C5, C32, C73, C87, C100, C111, C114, C129, C130, C141, C145, C146, C162, C163 +1
16100kResistor_SMD:R_0603_1608MetricR3, R8, R46, R56, R58, R82, R86, R94, R103, R120, R160, R178, R179, R181, R188, R189
14100@100MHz 3AInductor_SMD:L_0603_1608MetricFB3, FB4, FB10, FB11, FB12, FB14, FB15, FB16, FB17, FB18, FB19, FB20, FB21, FB22
114.7kResistor_SMD:R_0603_1608MetricR33, R43, R44, R68, R69, R131, R132, R135, R136, R164, R174
1010nFCapacitor_SMD:C_0603_1608MetricC52, C58, C102, C104, C116, C117, C128, C148, C152, C153
9100Resistor_SMD:R_0603_1608MetricR10, R11, R12, R13, R14, R15, R16, R175, R176
91MResistor_SMD:R_0603_1608MetricR4, R5, R6, R55, R61, R96, R97, R153, R156
81.5kResistor_SMD:R_0603_1608MetricR35, R37, R38, R40, R42, R71, R122, R195
815Resistor_SMD:R_2512_6332MetricR25, R26, R27, R28, R29, R30, R31, R32
83.3kResistor_SMD:R_0603_1608MetricR17, R18, R19, R20, R21, R22, R23, R24
833Resistor_SMD:R_0603_1608MetricR70, R72, R114, R121, R138, R139, R150, R151
8PMEG60T10ELRDiode_SMD:Nexperia_CFP3_SOD-123WD17, D18, D19, D24, D25, D26, D27, D28
8PMV50EPEARPackage_TO_SOT_SMD:SOT-23Q5, Q6, Q7, Q8, Q9, Q10, Q11, Q12
7220@100MHz 1.4AInductor_SMD:L_0603_1608MetricFB1, FB2, FB5, FB6, FB7, FB8, FB9
747kResistor_SMD:R_0603_1608MetricR47, R48, R80, R142, R143, R144, R197
618pFCapacitor_SMD:C_0603_1608MetricC8, C10, C48, C51, C62, C98
64.7uFCapacitor_SMD:C_0603_1608MetricC6, C33, C35, C139, C151, C170
547uFCapacitor_SMD:C_0603_1608MetricC83, C121, C123, C124, C126
415uHInductor_SMD:L_1210_3225MetricL8, L9, L10, L11
41kResistor_SMD:R_0603_1608MetricR67, R110, R183, R184
42.2uFCapacitor_SMD:C_0603_1608MetricC44, C82, C125, C160
427Resistor_SMD:R_0603_1608MetricR90, R91, R92, R93
42nFCapacitor_SMD:C_0603_1608MetricC64, C68, C78, C79
449.9Resistor_SMD:R_0603_1608MetricR123, R148, R198, R199
4MountingHole_Padfootprints:WA-SMSI-M2H22, H24, H25, H27
4MountingHole_PadMountingHole:MountingHole_2.2mm_M2_DIN965_PadH5, H6, H7, H8
4Si7850DPfootprints:PowerPAK_SO-8_Single_8PadsQ1, Q2, Q15, Q17
30157004.DRfootprints:fuse_smfF1, F2, F3
351kResistor_SMD:R_0603_1608MetricR53, R57, R64
3AP22815footprints:TSOT-25U20, U24, U27
3BAT46WJ,115Diode_SMD:D_SOD-323D2, D29, D30
3SBR0560S1Diode_SMD:D_SOD-123D3, D4, D6
3SP3012-06UTGfootprints:SON50P350X135X55-14NU21, U25, U26
3USB3_Afootprints:692121030100J5, J6, J17
20.01Resistor_SMD:R_0612_1632MetricR49, R50
20.02Resistor_SMD:R_0612_1632MetricR41, R177
20.33uFCapacitor_SMD:C_0603_1608MetricC3, C31
2100uFCapacitor_SMD:CP_Elec_8x10C16, C17
2100uFCapacitor_SMD:CP_Elec_6.3x7.7C122, C161
2150kResistor_SMD:R_0603_1608MetricR34, R180
22.2uHInductor_SMD:L_1210_3225MetricL6, L7
2220Resistor_SMD:R_0603_1608MetricR66, R112
2220uFCapacitor_SMD:C_1210_3225MetricC127, C147
2220uFCapacitor_SMD:CP_Elec_8x10C18, C19
2470nFCapacitor_SMD:C_0603_1608MetricC157, C159
26.8pFCapacitor_SMD:C_0603_1608MetricC36, C38
268uHInductor_SMD:L_Bourns_SRR1210AL3, L4
2BZX584C3V3-G3-08Diode_SMD:D_SOD-523D15, D16
2SK54A-LTPDiode_SMD:D_SMAD20, D21
2SW_DIP_x01Button_Switch_SMD:SW_DIP_SPSTx01_Slide_Omron_A6S-110x_W8.9mm_P2.54mmSW1, SW3
2SW_Pushfootprints:UK-B0206-G3.8-250-JZSW2, SW6
2Si7461DPfootprints:PowerPAK_SO-8_Single_8PadsQ3, Q18
2TLV62568DBVPackage_TO_SOT_SMD:SOT-23-5U13, U19
10.033uFCapacitor_SMD:C_0603_1608MetricC45
11-2199230-6footprints:TE_1-2199230-6J10
1100k NTCResistor_SMD:R_0603_1608MetricR79
110uHInductor_SMD:L_1210_3225MetricL5
112MHzCrystal:Crystal_SMD_Abracon_ABM8G-4Pin_3.2x2.5mmY2
122uH 5AInductor_SMD:L_12x12mm_H6mmL1
124MHzCrystal:Crystal_SMD_Abracon_ABM8G-4Pin_3.2x2.5mmY1
125MHzCrystal:Crystal_SMD_Abracon_ABM8G-4Pin_3.2x2.5mmY3
127kResistor_SMD:R_0603_1608MetricR192
132.768 kHzCrystal:Crystal_SMD_3215-2Pin_3.2x1.5mmY4
1475Resistor_SMD:R_0603_1608MetricR36
148099-5701footprints:MOLEX_48099-5701J11
149.9kResistor_SMD:R_0603_1608MetricR173
162Resistor_SMD:R_0603_1608MetricR167
1679105700footprints:MOLEX_67910-5700U11
1680Resistor_SMD:R_0603_1608MetricR140
1680pFCapacitor_SMD:C_0603_1608MetricC65
19.09K 1%Resistor_SMD:R_0603_1608MetricR59
190.9K 1%Resistor_SMD:R_0603_1608MetricR60
1BADGEfootprints:reform2-motherboard-badgeH14
1BATCON1Connector_Molex:Molex_Pico-Lock_504050-0591_1x05-1MP_P1.50mm_HorizontalJ13
1BATCON2Connector_Molex:Molex_Pico-Lock_504050-0591_1x05-1MP_P1.50mm_HorizontalJ2
1BZT52C10Diode_SMD:D_SOD-123D1

The 442 most common of 505 parts.

Similar boards

  • 6 layers276 × 93.5 mmReform2-motherboard30, the printed circuit board in 3D

    Reform2-motherboard30

    panyam/agni-samples

    This is the main repository for the MNT Reform open source laptop.

    529 partsApache-2.0
  • 4 layers140.8 × 32.1 mmPCB v1, the printed circuit board in 3D

    PCB v1

    EPFLRocketTeam/SupersonicTestRocket

    Repository for the avionics going on the supersonic test rocket

    75 parts4
  • 4 layers194.4 × 114.3 mmPocket reform keyboard kailh ortho, the printed circuit board in 3D

    Pocket reform keyboard kailh ortho

    torlye/pocket-reform

    My muscle memory requires me to have the Ctrl key at the bottom left corner of the keyboard!

    309 parts1
  • 4 layers32.9 × 50.2 mmVisual ADC Relay Block, the printed circuit board in 3D

    Visual ADC Relay Block

    baldengineer/VADC-visualized-adc

    Building an SAR ADC from scratch. Circuit schematics, PCB design, KiCad files, and Arduino code.

    34 parts3MIT
  • 2 layers105 × 105 mmGW A, the printed circuit board in 3D

    GW A

    ayaxan7/SolarAlert

    SolarAlert is a compact embedded hardware project designed to monitor the health and performance of solar panels.

    97 parts
  • 6 layers90.2 × 95.9 mmBoard sierra, the printed circuit board in 3D

    Board sierra

    visionspacetec/VST104-Sierra

    VST104 single OBC board - referred as board_sierra

    245 parts10

Questions

How big is the Mnt reform motherboard PCB?

It measures 276 × 93.5 mm and has 6 copper layers, 1.6 mm thick.

What parts does the Mnt reform motherboard use?

505 parts, built around the TLV62568DBV, AP22815AWT, SP3012-06UTG and PMEG60T10ELRX. The full list of its parts is on this page, with links to each part's symbol and footprint where PCBKit's library has them.

Where can I download the Mnt reform motherboard KiCad files?

From the Junkz3/wrench-board repository on GitHub, at the commit this page was made from: the board is web/boards/mnt-reform-motherboard.kicad_pcb. PCBKit opens the board in the browser, with nothing to install.

Can I use the Mnt reform motherboard design in my own project?

Its repository states no licence, so its author keeps all rights: ask Junkz3 before reusing the design.