Thumbdeck right

Gripster

A wireless thumb keyboard that turns your phone into a handheld deck. The phone clamps landscape in the middle on a spring-loaded expanding bridge (Kishi/Backbone style, 130–170 mm phones); two grips flank it, and you thumb-type a 78-key QWERTY on metal snap domes over Bluetooth.

_{Click the image to spin the assembly in 3D, pull it apart, and slide the clamp through its travel.}

How it works

The phone sits down inside a well so the glass and the keyboard face form one flat plane 14.7 mm off the back. Two contoured grips flank it, GBC-styled: boxy outline, translucent Atomic-Purple shell, dark-gray keymats.

Despite the two halves, this is not a split keyboard. One nRF52840 module in the right grip scans a single 9×10 matrix; the left grip has no active silicon. A single 20-way FFC ribbon crosses the spine carrying both the matrix and the battery, so there is one BLE device, one battery and one firmware image.

  LEFT grip (passive)                        RIGHT grip (MCU)
  42 domes + diodes  ── 20-way FFC jumper ──  Ebyte E73 (nRF52840)  ⇄ phone (BLE HID)
  D-pad/OK, mouse       14 matrix + VBAT      + charger + USB-C + power switch
  LiPo + J4 + F1 PTC    + GND, one cable      scans the whole 9×10 matrix

There are no switches and no hand-soldering: 78 Snaptron 7 mm snap domes press directly onto gold ENIG pads, and everything soldered is SMT on one side of each board, so JLCPCB builds it turnkey.

Everything in this repo — the KiCad boards, the printable shells, the renders, the ZMK keymap — is generated from one parametric Python file, hardware/scripts/deck.py. Move a key and the pad, the shell opening, the keycap and the firmware matrix all move with it.

Naming: the project is Gripster, but every internal identifier — the ZMK board id, the board and firmware filenames, the Bluetooth device name — is thumbdeck. Those are real strings in real files.

Specs

Building one

Full step-by-step in docs/assembly.md.

1. Order the boards. Regenerate the fab package (python3 hardware/scripts/gen_fab.py), then upload fab/right/ and fab/left/ as two separate JLCPCB PCBA orders. Options: 4-layer, 1.6 mm FR-4, ENIG, assembly side = bottom. The right board needs Standard assembly (the E73 is Extended); the left can go Economic. Check the DFM preview for LED polarity and SOT-23/USB-C/E73 rotation before paying. E73 stock is volatile — reserve C356849 before anything else.

2. Print and press. Shells in PETG, keymats in TPU 95A. Fatigue-test a 3×3 keymat coupon past 10 k cycles before committing a full mat. Press the 78 domes onto their pads under polyimide retention tape — the keymat alone will not stop a dome walking off its ring.

3. Assemble. FFC jumper first, before the lids go on: 20-way, type-A (contacts on the same side at both ends), contacts facing the board at both ends. Foam-tape the cell to the left grip floor before its board goes in, and meter the pigtail against the "+"/"−" silk at J4 on the left board — vendors wire PH pigtails both ways. Thread the ribbon through its walled tray lane before lapping the shrouds, press the TPU grippers onto the cradles, then the lids over them; 5× M2×12 per grip, driven from the back into the lid's inserts.

4. Flash. The E73 ships blank. One-time SWD flash of the Adafruit nRF52 bootloader (nice_nano build) via TP1–5, possibly after nrfjprog --recover. This needs a debug probe and is the most likely place to get stuck. After that it's drag-and-drop UF2 forever. Pair to the device advertising as thumbdeck.

Things that will bite you

• Do not fit a 16-way ribbon. It floats in the 20-way housing with 4 mm of slop at each end.
• Do not connect the cell before the first flash — the bootloader sets REGOUT0 to 3.3 V.
• Ribbon before cell, always. Two of its conductors are the raw cell, so seating it with the battery plugged in is a live operation. The switch gates the load, not the cell; J4 is the only way to de-energize the ribbon.
• The cell must have an integrated PCM. It and F1 cover different fault bands and neither substitutes for the other.
• HASL will not work. Snap domes need gold; HASL oxidises within weeks of cycling.
• Debugging: a dead column on the left points at the ribbon; a dead row kills that row on both grips.

Will it fit my phone?

The clamp takes any 130–170 mm long edge out of the box. The interactive viewer will seat a Galaxy S25 Ultra, an iPhone 17, an iPhone Air or an iPhone 17 Pro Max in it and re-span the jaw to each — the Phone panel in the control bar. The well itself is dimensioned for an S25 Ultra in a ~1.2 mm case; phone_w, phone_h, phone_t and case_t in deck.py re-generate the shells for something else, and the PCBs are unaffected. A substantially larger phone pushes the grips further apart than is comfortable to thumb-type across, and moves the bridge connectors, so re-check the ribbon length.

Repo map

Reproduce the design

cd hardware/scripts
python3 gen_board.py                  # placement + netlist + USB copper + GND escape vias
./route.sh right && ./route.sh left   # Freerouting -> GND stitch + zone fill -> DRC gate
python3 gen_fab.py                    # gerbers/BOM/CPL per side (refuses unless DRC-clean)
python3 verify_alignment.py           # 2D stack audit: pads vs model, gutters, boss clearances
python3 sim_matrix.py                 # ghosting/NKRO proof
python3 gen_firmware.py               # ZMK transform/keymap/gpio
python3 layout_gen.py --iter 23       # layout diagram
python3 render_fab.py                 # fab view
# --- 3D (CadQuery) ---
cd ../.. && hardware/cad/.venv/bin/python hardware/cad/deck3d.py --all --check --render
hardware/cad/.venv/bin/python hardware/cad/export_full_asm.py   # full-assembly GLB

Needs Python 3 + matplotlib; the board pipeline needs KiCad 9 (pcbnew + kicad-cli) and a Freerouting jar; the 3D generator uses a venv at hardware/cad/.venv/ built from hardware/cad/requirements.txt.

Routing intermediates, DRC JSONs and fab/ are build artifacts and are not committed. Generated files are outputs, not sources — don't hand-edit hardware/kicad/generated/, hardware/cad/models/ or renders/; change the model and re-run.

Open questions

Real decisions still on the table: a flat ortho grid vs a canted/fanned key arc; shoulder buttons for the index fingers; an I²C trackpad on the TP6–8 breakout (dropped from v1 over driver burden). A trackpad and shoulder buttons are both rev-B changes.

Contributing

See CONTRIBUTING.md and CODE_OF_CONDUCT.md. The most valuable contribution right now is not code — it's a first-article build reported with photos, especially if it doesn't work. After that: ergonomic critique, a KiCad/RF review of the autorouted boards, and a ZMK review of the board definition.

There is no kit, no store and no waiting list. It's Apache 2.0, so build it, modify it or sell it yourself.

License

Apache License 2.0 — see LICENSE. Copyright © 2026 Eric Neuman.

Some files carry their own terms and are not relicensed: the two marbastlib-derived footprints (CERN-OHL-P-2.0), stock KiCad footprints (CC-BY-SA-4.0 with the library exception), the phone GLB and the references/ photos (provenance not established — establish it or remove them before redistributing). Details in NOTICE and docs/reference-notes.md.

"Rii" is a trademark of its owner; this project is not affiliated with or endorsed by Rii. No i8+ artwork, geometry, footprint or firmware was copied into this design.