Taprivo

Power your code with motion. ⚡

Türkçe README

Taprivo turns hand movement into a playful work budget for AI coding agents. Squeeze your hands in front of the camera (or tap keys in the simulator), build Motion Energy, and let Claude Code read and spend it through MCP. Hand tracking runs locally on your device; camera frames never leave it. Motion Energy is a game mechanic, not API tokens or credits.

Status: beta (0.1.0b6). Camera-based squeeze detection with calibration is available alongside the keyboard simulator. Tested on macOS on Apple Silicon with the built-in FaceTime camera.

What it does

• Every tap adds 10 Motion Energy (× combo multiplier, up to 2×) to a session balance (cap 10,000); a camera squeeze counts as five taps (50 energy). Hold a steady beat and the energy per tap rises by another 25 %.
• A small always-on-top HUD (dark or light theme) leads with the energy number and its bar, then a hand map: two hand silhouettes whose finger tips light up as you drum, with the key caps under each finger. Combo and multiplier, rate and tempo, and camera and MCP status badges sit alongside.
• A local MCP server on http://127.0.0.1:32145/mcp exposes get_energy, spend_energy, get_stats and get_session.
• Claude Code reads the balance and spends a suitable amount before a substantial implementation burst, following an instruction file you opt into.
• Every Claude Code project on your machine shares one balance.

How Motion Energy works

available = generated - overflow - spent

generated counts every tap, overflow is energy lost while the balance is full, and spent is what agents charged. Spending is atomic: two clients can never spend the same energy, and retrying a spend with the same request id returns the original result instead of charging twice.

Energy is not a token balance, API credit or permission. Your existing task and agent permissions stay exactly as they are; running out of energy never blocks you, and the budget can be disabled at any time.

The balance resets when Taprivo quits, but the session is not forgotten: each session and its daily totals are written to a local SQLite file, so taprivo stats --today and taprivo stats --history still work with the app closed.

Requirements

• macOS on Apple Silicon (other platforms are untested)
• Python 3.12 or 3.13 and uv
• Claude Code for the agent integration

Install

Pick the path that fits you; the app is the easiest.

1. macOS app (easiest, Apple Silicon)

1. Download Taprivo-<version>.dmg from the latest release.

2. Open the DMG and drag Taprivo to Applications. The app is signed with a Developer ID and notarized by Apple, so it opens without extra steps.

3. Open Taprivo. The HUD appears; drum the number row (1–4 left hand, 7–0 right hand) and watch the energy grow, or click Open Camera to count hand squeezes. macOS asks for camera permission the first time.

4. Click Setup… in the HUD to connect Claude Code or Cursor and to run the doctor; the same window shows the command-line tool bundled in the app. The same thing from a Terminal:

   /Applications/Taprivo.app/Contents/MacOS/Taprivo setup claude --install-instructions
   /Applications/Taprivo.app/Contents/MacOS/Taprivo doctor
   

   Use setup cursor --install-instructions for Cursor. Keep the app running while you work; the balance resets when it quits.

2. Command line (uv or pipx)

uv tool install taprivo            # or: pipx install taprivo
taprivo                            # opens the HUD
taprivo setup claude --install-instructions
taprivo doctor

Needs Python 3.12 or 3.13. The install takes about 1.5 GB because of mediapipe, OpenCV and Qt. Upgrade with uv tool upgrade taprivo, remove with uv tool uninstall taprivo.

3. From source (contributors)

Follow the Quickstart below. Contributors can also build the app bundle with packaging/macos/build.sh build.

Quickstart (from a clone)

git clone https://github.com/yorulmazsinan/taprivo.git
cd taprivo
uv sync
uv run taprivo simulate

The HUD opens with keyboard mode running and both hands on the number row. The left hand drums 1 2 3 4 (pinky to index) and the right hand 7 8 9 0 (index to pinky). Each press adds 10 energy; keep the taps coming and the combo multiplier lifts that to 1.5× at 10 consecutive taps and 2× at 25. Keep an even beat between 60 and 240 BPM and the HUD highlights the tempo and adds a further 1.25×. Taps above 12 per second are ignored, so a held key earns nothing.

With a camera: click Open Camera in the HUD or run uv run taprivo calibrate — see Camera.

Connect Claude Code

uv run taprivo setup claude --install-instructions
uv run taprivo doctor

setup claude registers a user-scope HTTP MCP server named taprivo using the official claude mcp add command, writes ~/.claude/taprivo.md, and, with --install-instructions, imports that file from ~/.claude/CLAUDE.md inside a clearly marked block (a backup is written first). Run with --dry-run to see the changes without writing anything. doctor checks the endpoint, token, registration and instruction files.

To share the configuration inside a repository, run uv run taprivo setup claude --project. That writes a .mcp.json entry whose token comes from the TAPRIVO_TOKEN environment variable, so no secret is committed.

uv run taprivo remove claude undoes the registration and removes only the files and blocks Taprivo added.

Cursor

uv run taprivo setup cursor --install-instructions
uv run taprivo doctor

Use taprivo setup cursor --install-instructions when Taprivo is installed globally. setup cursor adds a taprivo entry to ~/.cursor/mcp.json (mode 0600, a backup is written first) and, with --install-instructions, writes ~/.cursor/taprivo.md. Cursor has no command for user rules, so paste the contents of that file into Settings > Rules yourself.

With --project it also writes .cursor/mcp.json in the repository, where the token comes from the ${env:TAPRIVO_TOKEN} environment variable, and .cursor/rules/taprivo.mdc with the same budget rules. --dry-run shows the changes without writing, and uv run taprivo remove cursor removes the entry and the files Taprivo added.

Camera

Taprivo tracks up to two hands in front of the camera and counts a squeeze each time a hand goes open → fist → open. Each squeeze is worth 50 Motion Energy. Think of it as a short circulation exercise between coding bursts.

1. Start Taprivo and click Open Camera in the HUD (or run uv run taprivo calibrate).
2. Pick a device. Taprivo prefers the built-in camera; an iPhone Continuity Camera is listed but not opened until you select it (set camera.prefer_builtin: false to change the default).
3. Click Start Camera. macOS asks for camera permission the first time.
4. Click Calibrate and follow the prompts: show your hand, open it wide, make a fist, then squeeze five times. Apply the result for this session.

Closing the Camera window stops the camera; energy from squeezes stays in the HUD.

The Camera window draws the tracked hand over the live preview — 21 landmarks with per-finger fingertip colours, and an expanding ring from the palm each time a squeeze is counted — plus a strip with the frame rate, the detection ratio and the device name. It also shows a Left/Right openness meter per hand with the calibrated open and closed levels as tick marks and the current state beside it. Set hud.reduced_motion: true for a short static ring instead of the expanding one. Camera frames are processed in memory and shown only in the Camera window; they are never stored, logged or exposed through MCP. Calibration data can be exported as a CSV of per-hand features (no images, no raw landmarks) for tuning.

Known limitations: good lighting and the whole hand in frame are needed; very quick squeezes are ignored; typing with your hands in view can occasionally be counted. Per-finger tapping is not detected by the camera — use the keyboard simulator for that. taprivo doctor --camera-probe opens the camera to check permission and report processed fps. A MacBook with its lid closed keeps the built-in camera dark; open the lid or pick another camera. Keep the whole hand inside the frame; a hand partly out of view is ignored.

CLI

Exit codes: 0 success, 1 operation failure, 2 invalid arguments or config.

Configuration

User overrides live in ~/.config/taprivo/config.yaml and merge over the shipped defaults (src/taprivo/resources/default.yaml). Common keys:

energy:
  energy_per_tap: 10
  max_energy: 10000
combo:
  energy_multiplier_enabled: true
  tiers:            # combo tiers: consecutive taps → energy multiplier
    - {at: 10, multiplier: 1.5}
    - {at: 25, multiplier: 2.0}
rhythm:
  enabled: true
  steady_multiplier: 1.25   # steady beat bonus, stacks with combo
server:
  port: 32145
simulator:
  max_taps_per_second: 12   # sliding one-second cap, both hands together
hud:
  always_on_top: true
  opacity: 0.92
  reduced_motion: false
  theme: system   # system | dark | light
camera:
  device_index: null   # null = choose in the Camera window
  prefer_builtin: true # false = pick the first camera with a signal instead
  width: 640
  height: 480
squeeze:
  open_level: 0.80     # calibration overrides both levels for the session
  closed_level: 0.45
stats:
  enabled: true        # session and daily totals; no reasons, no frames
  path: null           # null = ~/.config/taprivo/stats.sqlite

If port 32145 is taken, the HUD shows MCP: Error; set server.port and run taprivo setup claude again. Taprivo never switches ports silently.

Privacy and security

• The MCP server binds to loopback only, checks Host and Origin, and requires a random per-user bearer token stored at ~/.config/taprivo/token (mode 0600). No CORS headers are sent.
• Camera frames are processed in memory and are never stored, uploaded or exposed over MCP. The hand-tracking model runs on-device; mediapipe is pinned to a release without usage logging.
• Statistics are stored locally in ~/.config/taprivo/stats.sqlite as counters and timestamps only — no spend reasons, no camera data; disable with stats.enabled: false.
• No telemetry. The only network listener is the local endpoint.
• Logs stay low-volume and never contain the token or spend reasons in full.

Architecture

Camera (MediaPipe hand landmarks, squeeze detector) / keyboard simulator
  → TapEvent
  → EnergyEngine (single lock, atomic spend, idempotency)
      ├─ HUD (PySide6)
      └─ MCP server (Streamable HTTP, loopback, bearer token)
           ├─ Claude Code project A
           └─ Claude Code project B

Support matrix

See ROADMAP.md.

Contributing

Start with the simulator; no camera is needed to contribute. Read CONTRIBUTING.md, pick a good first issue, and open a PR.

License

Apache-2.0. See LICENSE and THIRD_PARTY_NOTICES.md.