Skip to content
clevercat64Public

About

Your WhatsApp, answered by your own AI agents over MCP. A CleverCoder plugin for Windows: you decide, chat by chat, how much they're allowed to do — from answering on their own to nothing leaving without your approval. Runs on your machine, open-source.

Topics

Resources

Stars

15 stars

Watchers

1 watching

Forks

Latest commit

 

History

19 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

piumy-gateway

The engine behind Piumy: your WhatsApp, answered by your own AI agents over MCP — routed, stored, and under your rules.

Everything runs on your machine. There are no servers in between, because there are none.

Getting it

You do not need this repository to use Piumy. Download the installer — one .exe, one double click, Windows only for now.

This repository is the source: for reading it, auditing it, or building it yourself.

What it is

A single Go binary that sits between WhatsApp and your agents:

  • WhatsApp — whatsmeow embedded in-process. No browser, no Node, no external client. CGO_ENABLED=0.
  • Agents — an MCP server they connect to, plus a dispatch channel that pushes incoming messages to them.
  • Storage — SQLite (pure Go), holding history, per-chat rules, memory and the outbox.
  • A dashboard on localhost to see and steer all of it.

An agent cannot act on its own: every action must belong to a real incoming message it is currently handling. That gate lives in the code, not in a prompt.

Building

go build ./...
go test ./...

The Windows installer is built from installer/windows/ with Inno Setup.

Layout

Path What
internal/whatsmeow the WhatsApp client adapter
internal/corepipeline what happens to a message from arrival to answer
internal/mcpserver the tools agents call, and the gate that limits them
internal/capipush dispatching messages out to agents
internal/store SQLite: chats, messages, rules, memory, drafts
internal/dashboard, internal/restapi the local dashboard and its API
docs/MANUAL.md the map: every package and its exported surface

Agent-facing manuals live in internal/mcpserver/manuals/ and are embedded in the binary — an agent reads them over MCP with get_manual, so they cannot drift from the build.

Contributing

Two rules carry more weight than the rest:

  • No personal data in the repository — not even test data. No phone numbers, WhatsApp identifiers, emails or real names; no chat, contact or history dumps. When a test needs an identifier, use the 555 prefix (555000001@s.whatsapp.net). A pre-commit hook enforces this: enable it with git config core.hooksPath .githooks.
  • The hard limits belong in the code, never in a prompt or a manual. A rule an agent can be talked out of is not a rule.

License

GNU Affero General Public License v3.0 — see LICENSE.

About

Your WhatsApp, answered by your own AI agents over MCP. A CleverCoder plugin for Windows: you decide, chat by chat, how much they're allowed to do — from answering on their own to nothing leaving without your approval. Runs on your machine, open-source.

Topics

Resources

Stars

15 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages