Skip to content
Shlok Sawant
← all work

Ribhus

An open-source WhatsApp agent

Role
AI Engineer
Year
2026 · open source
Links
github ↗

Stack

  • Python
  • TypeScript
  • Playwright
  • SQLite
  • OpenAI
  • MCP

An open-source agent that runs on your machine. It searches, browses, remembers, and drafts. WhatsApp and Gmail go out only after you approve them. You pick the model.

Problem

A useful agent has to look things up and then act on the answer. The risky part is the send: a message that leaves the machine without a person checking it. A hosted agent also keeps the memory and the session somewhere else.

Constraints

  • It runs locally. The model is OpenAI, OpenRouter, or a local one such as Ollama.
  • WhatsApp and Gmail are driven in Chromium through the websites, not the official APIs. The project does not promise account safety.
  • A send always waits for approve, edit, or cancel. Auto-approve cannot skip those. One recipient per send, with rate limits and a cooldown.
  • Memory sits in a local SQLite file. It can be listed, edited, exported, or turned off. Older episodes are pruned to about 80.
  • Python 3.11+ and uv. MIT license.

Architecture

A CLI, plus a local dashboard. Tools cover web search, page extract, a browser, files, Python, and shell. Skills add workflows: Indian company research, travel shortlists, and football briefings. MCP tools take the same approval path as the built-in ones.

Higher-risk actions ask first and land in an audit log. WhatsApp and Gmail are separate from the generic browser, which cannot open those sites. Session cookies stay in a local profile. The dashboard listens on 127.0.0.1:8742 for tasks, skills, memory, channels, MCP, and config.

Demo

The repo is the product: github.com/Shlok221B/Ribhus.