wactx

April 15, 2026 · View on GitHub

Semantic and graph search for your WhatsApp messages. All data stays local on your machine. No servers, no cloud, and no telemetry.

What it does: Sync your WhatsApp via whatsmeow, embed with any OpenAI-compatible API, build a relationship graph, and search across people, topics, and messages in one query.

Everything is contained in two files:

  • ~/.config/wactx/config.toml (configuration and session)
  • ~/.local/share/wactx/messages.duckdb (database)

Quickstart

Install wactx immediately via pip or uv. No Go compiler or build steps are required.

pip install whatsapp-ctx-cli

or

uv tool install whatsapp-ctx-cli

After installation, the wactx command is available in your PATH.

wactx init                    # interactive setup for provider and API key
wactx sync                    # scan QR code on first run
wactx index                   # embed messages
wactx search "who knows about fundraising"

wactx init walks you through setting up your API provider and key. It handles the setup automatically.

Disclaimer

This tool uses whatsmeow, which is an unofficial Go library for the WhatsApp Web multi-device API.

  • Usage is read-only. This tool only reads messages and never sends or modifies anything.
  • Using unofficial WhatsApp APIs carries inherent risk. WhatsApp could restrict or ban accounts that use third-party clients.
  • The authors are not responsible for any account actions taken by WhatsApp.
  • Use this tool at your own discretion. It is best to test it on non-critical accounts first.

Data and Privacy

Privacy is the core focus of this project.

  • Local Storage: All synced messages, contacts, and metadata are stored locally.
  • Database: Found at ~/.local/share/wactx/messages.duckdb.
  • Config: Found at ~/.config/wactx/config.toml.
  • WhatsApp Session: The whatsmeow.db session file is stored alongside your config.
  • Removal: To remove all data, run wactx clean or manually delete the files listed above.
  • Embeddings: Message embeddings are sent to the API provider you configure, such as OpenAI or Cloudflare.
  • Zero Data Leak: If you use a local Ollama instance for embeddings, no data ever leaves your machine.
wactx search "kubernetes expert" --depth fast        # ~2s, 1 query, no graph
wactx search "fundraising advice"                    # ~5s, balanced (default)
wactx search "AI research" --depth deep              # ~8s, 8 query variants, full graph

wactx search "cofounder" --variants 3 --top 20
wactx search "sales strategy" --no-graph
wactx search "hiring ML engineers" --json
wactx search "investors" --json | jq '.people[:5]'

Filters

Narrow results with literal-keyword, chat, and time-range filters. Filters apply to BM25, semantic, and graph-expansion passes in one place.

wactx search "expert" -k Kubernetes                  # must contain "Kubernetes"
wactx search "expert" -k Kubernetes -k Google        # must contain BOTH (AND)
wactx search "advice" --chat "Founders Chat"         # substring-match chat name
wactx search "advice" --chat 120363@g.us             # or a JID directly
wactx search "kubernetes" --after 2026-01-01 --before 2026-03-31
wactx search "production" -k Docker --chat "Founders" --after 2026-01-01
  • -k, --keyword TEXT — repeatable; each one is a required case-insensitive substring (AND).
  • --chat TEXT — either a JID (...@g.us / ...@s.whatsapp.net) or a case-insensitive substring matched against contact/group names. Fails fast with a clear error if nothing matches.
  • --after YYYY-MM-DD / --before YYYY-MM-DD — inclusive date range on the message timestamp. --before 2026-03-31 includes all of March 31.

Search returns two ranked lists:

  • People: Scored by semantic similarity and graph proximity. This includes DMs, shared groups, and co-mentioned entities.
  • Messages: Ranked by embedding cosine similarity across multiple query variants.

When graph data is available, you also get a Graph Insights panel. This shows shared groups between people, common entities, and your connection strength (🟢 strong, 🟡 weak, or ⚪ indirect).

Sync and Media

wactx sync                    # incremental sync (default)
wactx sync --full             # full history sync
wactx sync --live             # stay connected to receive new messages

wactx download                # download all media
wactx download --types image --after 2026-01-01
wactx download --chat "group@g.us"

The first run displays a QR code in your terminal. Scan it with the WhatsApp app on your phone to connect. Your session persists across runs.

Entity Extraction and Graph

wactx enrich                  # extract persons, orgs, techs, URLs, and events
wactx graph                   # build relationship graph
wactx search "who should I talk to about fundraising"

The graph connects:

  • People to People via DMs and group co-membership.
  • Group memberships via message activity.
  • Entity mentions via extracted persons, organizations, technologies, and events.

Configuration

# ~/.config/wactx/config.toml

db_path = "~/.local/share/wactx/messages.duckdb"

[api]
base_url = "https://api.openai.com/v1"
key = ""
embedding_model = "text-embedding-3-large"
embedding_dims = 384
chat_model = "gpt-5-mini"
max_concurrent = 5

[sync]
wa_db_path = "whatsmeow.db"
media_dir = "media"
timeout = "5m"

[search]
default_depth = "balanced"
owner_name = "Your Name"

The tool works with any OpenAI-compatible endpoint:

# Cloudflare AI Gateway
wactx config api.base_url https://gateway.ai.cloudflare.com/v1/ACCOUNT/GATEWAY/compat

# Ollama (local and free)
wactx config api.base_url http://localhost:11434/v1
wactx config api.embedding_model nomic-embed-text

All Commands

CommandDescription
wactx initCreate config and database
wactx config KEY VALUESet a config value
wactx syncSync messages from WhatsApp
wactx downloadDownload media attachments
wactx index [--reset]Embed messages for semantic search
wactx enrich [--all]Extract entities from messages
wactx graphBuild relationship graph
wactx search QUERY [-k KW] [--chat ...] [--after D] [--before D] [--json]Semantic and graph search, with optional filters
wactx stats [--json]Show database statistics

Agent Integration

wactx is built to be agent-friendly: JSON output on search and stats, non-zero exit codes on error, structured filters, and no interactive prompts on any read-only command. A bundled skill teaches Claude Code how to drive it.

cp -r skills/wactx ~/.claude/skills/      # global install

Or, if you only want the skill inside this repo, Claude Code picks up skills/wactx/SKILL.md automatically when you run it from the project root.

Once installed, Claude will call wactx search --json (with the right filters) whenever you ask about contacts, conversations, or relationships. See skills/wactx/SKILL.md for the full contract — input flags, JSON schema, exit codes, and common recipes.

Architecture

WhatsApp (phone)
    │
    ▼
wactx sync (bundled Go binary, whatsmeow) ──→ messages + contacts (DuckDB)
    │
    ▼
wactx index  ──→ embeddings via OpenAI-compatible API + HNSW index
    │
    ▼
wactx enrich ──→ extracted entities (persons, orgs, techs, events)
    │
    ▼
wactx graph  ──→ DuckPGQ property graph (vertices + edges)
    │
    ▼
wactx search ──→ multi-query semantic search + graph traversal + rich output

Stack: Python, Go, DuckDB, DuckPGQ, DuckDB VSS, whatsmeow, OpenAI-compatible API, Rich, and Click.

Development

This section is for contributors. It requires uv and Go 1.21+.

git clone https://github.com/iofold/whatsapp-ctx-cli
cd whatsapp-ctx-cli
uv sync --group dev                     # install dependencies and dev tools
uv run python build_go.py               # compile Go binary
uv run pytest                           # run tests

Cross-compile for all platforms:

uv run python build_go.py --all         # linux/amd64, linux/arm64, darwin/amd64, darwin/arm64, windows/amd64

A Makefile is included as a convenience, but uv is all you need.

License

MIT