Commit

Commit enhances your WhatsApp usage by finding commitments automatically — things you said you'd do, and things others said they'd do. It auto detects when things are done.

When something goes quiet, it surfaces it for follow-up.

How it works

You scan a QR code to link your WhatsApp account, same as WhatsApp Web. Commit runs on your machine. Every 10 seconds it reads new messages, analyzes them with a local MLX Gemma 4 model, indexes them with EmbeddingGemma, and logs any commitments it finds.

The dashboard shows everything grouped by person: what you owe, what they owe, how long it's been. You can reply, set reminders, mark things done, or dismiss them.

Features

• Dashboard — all open commitments in one view, grouped by chat. Filter by "I owe", "They owe", or "Resolved". Search across everything.
• Auto-extraction — local Gemma 4 reads incoming messages every 10 seconds, identifies promises and obligations, and logs them with the original quote, person, and direction.
• Semantic search — EmbeddingGemma indexes WhatsApp text locally so @find can use meaning-based retrieval in addition to keyword search.
• Local Gemma 4 — uses the working 12B MLX path with an MTP draft model for local extraction.
• Auto-resolution — when a commitment is fulfilled in conversation (a file shared, a task confirmed, someone says "done"), Commit marks it resolved automatically.
• Follow-ups — surfaces things others owe you that have gone quiet. Drafts a polite nudge message and lets you send it directly from the dashboard.
• Reminders — set a reminder on any commitment. When it's due, Commit sends you a WhatsApp message to your own account so it shows up in your chat list.
• Favorites — star important chats or commitments to pin them to a dedicated tab for quick access.
• Reply from dashboard — respond to any commitment's chat directly from the Commit UI without switching to WhatsApp.
• WhatsApp bot commands — message yourself on WhatsApp to check commitments, search, or mark things done (see commands below).
• History sync — when you link a new device, Commit backfills your recent WhatsApp message history so commitments appear immediately.
• Dark and light theme — toggle from the dashboard toolbar. Respects your preference across sessions.
• Mobile web — the dashboard is fully responsive. Access it from your phone's browser at the same local address.
• Passcode protection — the web interface is secured with a passcode. Local model settings are stored on disk.

System requirements

• macOS 12 Monterey or later on Apple Silicon
• MLX model serving with an OpenAI-compatible /v1/chat/completions endpoint
• Commit starts its own local OpenAI-compatible /v1/embeddings endpoint backed by EmbeddingGemma
• WhatsApp account with multi-device support

Install

Mac (DMG)

Download Commit-x.x.x.dmg from Releases, open it, drag Commit to Applications, and open Commit from Applications.

Windows

Download Commit-x.x.x-windows-amd64.zip from Releases, extract it, and run Commit.exe.

From source

Requires Go 1.22+ and local MLX model servers.

git clone https://github.com/mitensampat/commit.git
cd commit
go build -o commit .
./commit

Local MLX models

Commit defaults to:

• Generation: mlx-community/gemma-4-12B-it-qat-4bit
• MTP draft model: mlx-community/gemma-4-12B-it-qat-assistant-nvfp4
• Embeddings: mlx-community/embeddinggemma-300m-4bit
• Chat endpoint: http://127.0.0.1:8080/v1
• Embedding endpoint: http://127.0.0.1:8081/v1

Gemma 4 E2B/E4B checkpoints are disabled for now because the current MLX VLM loader rejects their audio-tower weights with a shape mismatch. Commit normalizes those selections back to the working 12B MTP path until upstream support is fixed.

Experimental smaller MLX VLM options:

• mlx-community/Qwen3-VL-2B-Instruct-4bit
• mlx-community/SmolVLM-256M-Instruct-4bit

These smaller models run without an MTP draft model.

On first run, Commit checks the standard Hugging Face cache and downloads these repos with hf download or huggingface-cli download if they are missing. Install the Hugging Face Hub CLI with pipx first:

brew install pipx
pipx install "huggingface-hub[hf_xet]"

Commit shows dependency/model download status in the setup screen, repairs the local Gemma runtime when needed, then starts mlx_vlm.server for chat on port 8080 and a bundled EmbeddingGemma server for semantic search on port 8081. The bundled scripts/start-mlx-gemma.sh remains available for debugging, but normal users should not need a terminal.

Environment overrides:

COMMIT_LLM_BASE_URL=http://127.0.0.1:8080/v1
COMMIT_EMBEDDING_BASE_URL=http://127.0.0.1:8081/v1
COMMIT_LLM_MODEL=mlx-community/gemma-4-12B-it-qat-4bit
COMMIT_LLM_DRAFT_MODEL=mlx-community/gemma-4-12B-it-qat-assistant-nvfp4
COMMIT_EMBEDDING_MODEL=mlx-community/embeddinggemma-300m-4bit

For higher-accuracy local inference with the original MTP setup:

COMMIT_LLM_MODEL=mlx-community/gemma-4-12B-it-qat-4bit
COMMIT_LLM_DRAFT_MODEL=mlx-community/gemma-4-12B-it-qat-assistant-nvfp4
COMMIT_LLM_NUM_DRAFT_TOKENS=3

Setup

Open http://commit:9384 in your browser (or localhost:9384). The setup wizard will walk you through:

1. Set a passcode — protects the web interface
2. Check Local Gemma — validates the local MLX generation endpoint
3. Scan the QR code — links Commit to your WhatsApp account

Once connected, Commit scans incoming messages every 10 seconds and populates the dashboard.

WhatsApp bot commands

Message yourself on WhatsApp to interact with Commit:

Command	Description
commitments	List all open commitments
owe @person	Show what you owe someone
done <text>	Mark a commitment as resolved
search <query>	Find commitments by keyword
help	Show available commands

Architecture

main.go              — entry point, wires up components
server/              — HTTP server, API endpoints, auth
server/static/       — embedded web UI (single-page app)
extraction/          — local MLX clients, semantic indexing, commitment extraction prompt
store/               — SQLite database, message and commitment storage
whatsapp/            — WhatsApp client (whatsmeow), bot command handler
landing/             — marketing landing page

Data is stored in ~/.commit/:

• commit.db — SQLite database (messages, commitments, settings)
• WhatsApp session files

Building releases

# Mac DMG (ad-hoc signed, for local testing)
./scripts/build-mac.sh

# Mac DMG (signed + notarized, for distribution)
DEVELOPER_ID="Developer ID Application: ..." \
NOTARY_PROFILE="commit-notary" \
./scripts/build-mac.sh

# Windows zip
./scripts/build-windows.sh

# Both platforms
./scripts/build-all.sh

Privacy & data

• All data stays on your machine in ~/.commit/ — messages, commitments, settings, and WhatsApp session files
• Messages are decrypted locally by the WhatsApp multi-device protocol
• Message content, sender names, timestamps, embeddings, commitments, and search context stay local when your MLX endpoints are local
• No cloud storage, no telemetry, no tracking by Commit
• Your WhatsApp linked device session persists until you unlink it from your phone (Settings → Linked Devices) — even if Commit is not running, messages will queue for the next session
• To fully remove Commit: unlink the device from WhatsApp, delete the app, and delete ~/.commit/

Third-party services

• MLX / Gemma 4 — local commitment extraction, media descriptions, and nudge generation
• EmbeddingGemma — local semantic indexing for WhatsApp search

License

Open Source. Fork away and play.