WhatsApp AgentKit workflow
Claude Code interviews you about your business, then writes, tests and deploys a WhatsApp bot with per-customer memory
- 551 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent builds a server that handles customer conversations and holds Anthropic and WhatsApp keys.
- An empty webhook secret disables signature checks, so fill it before launch.
Install
- Works in
Manual install
git clone https://github.com/Hainrixz/whatsapp-agentkit.gitThen `bash start.sh` and the `/build-agent` command in Claude Code inside the folder.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent builds a server that handles customer conversations and holds Anthropic and WhatsApp keys.
- An empty webhook secret disables signature checks, so fill it before launch.
This is third-party code. Review the repository files before installing.
What it does
WhatsApp AgentKit is a `CLAUDE.md` file, a `/build-agent` command and a `start.sh` script. After you clone the repository, Claude Code reads these instructions and goes through five phases. First it checks the environment: Python 3.11 or newer, folders, dependencies, `.env`. Then it asks ten questions: business name and field, the bot's purpose, its name, tone, working hours, price or menu files, an Anthropic key and the way to connect to WhatsApp. From your answers the agent writes the project: the message server `agent/main.py`, the Claude call in `brain.py`, per-customer history in `memory.py`, business tools in `tools.py`, a provider adapter, `config/business.yaml` and `config/prompts.yaml`. A Dockerfile and a chat simulator `tests/test_local.py` come with it, where you write as a customer. WhatsApp connects through Zernio or directly through the Meta Cloud API. The server answers the webhook with 200 at once and processes the message in the background, drops duplicates by event id and checks the HMAC-SHA256 signature. At the end Claude Code helps deploy the project to Railway.
Who it is for. Owners of small businesses and freelancers who need an AI assistant in WhatsApp to answer customer questions and prefer answering interview questions to programming.
Good fit when
- A cafe or shop needs a bot that answers questions about hours and prices from a menu file
- You need a bot prototype in half an hour to show to a business owner
- A reply to a customer must take their past conversations into account
- A Meta Cloud API connection is already set up and only the agent remains to be written
- Later you switch from Zernio to the Meta Cloud API: just tell Claude Code
Not a fit when
- A ready cloud service without deploying your own server: AgentKit runs on your hosting
- The bot must run without Claude: the project is tied to the Anthropic API
- A bot for a different messenger: the project is built around WhatsApp
- You cannot pay for Meta messages and an Anthropic key: only AgentKit itself is free
Example request
/build-agent. A bot for a coffee shop: answers questions about the menu and hours, speaks in a friendly tone and hands over to a person when it does not know.Limitations
- You need Python 3.11 or newer, Claude Code, an Anthropic API key and a WhatsApp API account: Zernio or the Meta Cloud API. AgentKit itself is free, while conversations are paid to Meta.
- If `ZERNIO_WEBHOOK_SECRET` or `META_APP_SECRET` stays empty, the agent starts but skips signature checks and lets any message through. Fill the key before real customers arrive.
- The interface, README and instructions are written in Spanish.
- The repository is a starter kit: Claude Code writes the final code for your business, and you need to review it.
How to disable. Delete the cloned folder and the generated project. To stop the bot, disable the webhook at the provider and the service on Railway.
FAQ
How do Zernio and the Meta Cloud API differ?
Zernio handles the WhatsApp Cloud API connection: no Facebook app and no business verification, and the first two accounts are free. The Meta Cloud API is direct: you need a Facebook Business app, app review and a verified business.
Do I need to code?
No, you answer Claude Code's questions and it writes the code. You should still check the result: add test dialogs to `tests/test_local.py`.
How do I test the bot before launch?
Phase 4 opens a terminal chat where you write as a customer. If an answer is off, tell Claude Code and it adjusts the prompt.
Why does the bot not answer a repeated message twice?
Providers deliver a message at least once and can retry up to seven times. The server answers 200 at once and drops duplicates by event id.
How much does the bot cost to run?
AgentKit is free under the MIT license. You pay Meta for conversations, Anthropic for Claude requests and, if you want, Zernio beyond the two free accounts.
Discussion
No questions or reviews yet.
Related
WorkflowsConnects Claude Code, Cursor, Codex or OpenCode to Telegram, Slack and Discord: teammates message a bot, an agent replies
A plugin gives Claude Code OpenBSP API access and WhatsApp replies, forwarding channel messages only from selected contacts.
WeChat Automation API
wechat-automation-api
A skill for agents on Windows that sends text, images and files through an already running desktop WeChat, with an HTTP API and a queue as well
A CLI that passes your WeChat messages to an agent through the official ClawBot (iLink) and sends replies back
An MCP channel that passes WeChat messages into a Claude Code session and sends the agent's replies back to the same chat
An open-core platform: workflows and actions of connected services are published as MCP tools for Claude, Cursor and Windsurf
README in short
The README covers a two-step start, five build phases, the tree of the generated project, a Zernio versus Meta Cloud API comparison, a message flow and a cost table. It also explains three design choices: reply first, deduplication and signature checks.