Local connector

Normally your AI only reads WhatsApp when it is already working. If you write to it while it sits idle, your message waits. The zapbip connector fixes that. It is a small program that stays open on your computer. When you write on WhatsApp, it wakes the AI session you were already using, hands it your message and lets it answer you. It takes about 7 seconds for Claude Code and 6 for Codex. It works with Claude Code and Codex. Your computer must be on and awake, and the connector must be running.

Commands

Command	What it does
`connect`	Wakes a closed session when you write. Takes `--client`, `--cwd` and `--session`.
`listen`	Prints one JSON line per message, for an agent that watches a background command.
`channel`	Runs a local MCP channel that puts each message into an open Claude Code session.

ZAPBIP_CONNECTION_URL holds the connection link. connect and listen ask for it when it is not set.

What you need

Claude Code or Codex, already connected to zapbip.
At least one earlier session in the folder you want the AI to work in.
Node 24 or newer.
Your connection link, the same https://.../c/<secret>/mcp link your AI uses to talk to zapbip. Treat it like a password.

Start it

Open a terminal in the folder where you use your AI and run one of these:

export ZAPBIP_CONNECTION_URL='https://<host>/c/<secret>/mcp'
npx zapbip@0.1.0 connect --client claude-code

export ZAPBIP_CONNECTION_URL='https://<host>/c/<secret>/mcp'
npx zapbip@0.1.0 connect --client codex

If you do not set ZAPBIP_CONNECTION_URL, the connector asks for the link. It never takes the link as an argument, because arguments show up in your shell history and in the process list. Leave the terminal open. The connector prints one line for every event and stops when you press Ctrl+C. To keep it running without a terminal, see Run it in the background.

What happens when you write

You write to your AI on WhatsApp.
The connector receives your message and wakes the newest session in the folder.
The AI reads your message and answers you on WhatsApp.
If several messages arrive together, they are handled one at a time, in order.

If you were offline or the connector was stopped, your messages are not lost. When it connects again, it picks up every message that was not answered.

Choose the session

By default the connector uses the newest session in the folder. If you have the same folder open in another window, a wake can add to that conversation. To avoid it, pin a session:

npx zapbip@0.1.0 connect --client claude-code --session <session-id>

Option	What it does
`--client`	`claude-code` or `codex`. Required.
`--cwd`	The folder to work in. The default is the folder you ran the command in.
`--session`	A session id (UUID) to always resume, instead of the newest one.

What the AI can do when woken

Only one thing: send a message back to you. The connector gives the AI a private configuration that holds only zapbip, so tools from your other MCP servers are not available to it. Claude Code also runs without your settings files and with every built-in tool off, and Codex runs with a read-only sandbox and with the other MCP servers it finds in your Codex configuration files turned off. If it cannot read one of those files, it does not wake Codex and leaves the message unread. It also refuses when a managed Codex file sets any key outside a short list of harmless ones, when a profile sets sandbox, approval, tools or features (or any key outside that list), when a file turns a web search switch on, or when zapbip is defined anywhere but your Codex home config.toml. Codex web search is turned off. What remains for Codex is send_message and a read-only shell; apps and connectors of your Codex account are not turned off, and the macOS managed preferences com.openai.codex and the managed policy settings of Claude Code still apply. Your message reaches it as a message from you, marked as verified. A message you forwarded from someone else arrives marked as not verified, with a forwarded label, and the AI does not treat it as your instruction. If the Codex app already has the conversation open, Codex refuses to resume it. The connector then starts one new session in the same folder, which starts without the earlier conversation, and uses it from then on.

Keep an open session listening

connect wakes a session that is closed. For a session you have open right now, Claude Code can hear you at once, with no wake and no waiting. There are two ways, and the first is the better one. The channel is the first choice. It is a small program that Claude Code starts itself, in the same way it starts any other tool. It puts each message you write on WhatsApp straight into the open session, as a message from you, marked as verified (a forwarded message is marked as not verified). Set ZAPBIP_CONNECTION_URL in the environment Claude Code runs in, and Claude Code starts it with:

npx zapbip@0.1.0 channel

The channel cannot ask you for the link, so it only reads it from ZAPBIP_CONNECTION_URL. It does not take arguments. Declare both servers in the .mcp.json of the folder where you work. The channel brings your messages in; the zapbip HTTP server, under another key, gives the session the send_message tool it replies with:

{
  "mcpServers": {
    "zapbip-channel": {
      "command": "npx",
      "args": ["zapbip@0.1.0", "channel"],
      "env": { "ZAPBIP_CONNECTION_URL": "https://<host>/c/<secret>/mcp" }
    },
    "zapbip": {
      "type": "http",
      "url": "https://<host>/c/<secret>/mcp"
    }
  }
}

The session answers through zapbip. Without it the session hears you but has no way to reply. Then start Claude Code with:

claude --dangerously-load-development-channels server:zapbip-channel

zapbip is the key of the channel server in .mcp.json. The startup line shows Channels (experimental). Claude Code asks you to accept the development channel and the project MCP server; accept both. The listener is the backup. It prints one line for every message you write, and Claude Code watches those lines with its Monitor tool. Each line wakes the open session, and nothing is spent while you are silent:

export ZAPBIP_CONNECTION_URL='https://<host>/c/<secret>/mcp'
npx zapbip@0.1.0 listen

Each line is JSON with the message id, its kind, "sender":"owner", "verified" (true for your own message, false for a forwarded one), "forwarded" and the text. A Monitor watch lasts up to 30 minutes, so the AI starts the watch again each time one ends. Other agents that can keep a background command running and react to its output, such as OpenClaw and Hermes, can use npx zapbip@0.1.0 listen the same way. Both programs stay open and reconnect by themselves, like connect. Both mark a message as handled only after they pass it on, so a message is never lost and never shown twice. Their logs go to the error output and never carry your messages or your link. So the three ways fit together: the channel for an open session, listen as its backup, and connect for a closed one.

One connector per connection

If you start a second connector for the same connection, the first one finishes the wake it is running and stops. Run it on one computer at a time. If a message arrives at the moment you switch, both connectors can wake your AI for that same message once; zapbip delivers at least once, never zero times. While a connector is attached, zapbip wakes your AI through it. When it closes, zapbip goes back to the setting you had before.

What it prints

Each line is JSON with a time and an event, such as wake_started and wake_finished. The lines carry ids, kinds, outcomes and timings. They never carry your messages, the AI’s replies or your connection link.

When something does not work

You see	What it means
The connector exits at once	The link is unknown or revoked. Ask for a new one.
Your message gets no answer	There is no session in the folder yet, or `claude` or `codex` is not installed where the connector runs. Open the AI once in that folder.
It keeps reconnecting	Your network is down. It retries with growing waits, up to 30 seconds, and picks up where it left off.
It stopped by itself	You started another connector for the same connection.

Get started

Agents

Features

Help

For developers

Local connector

Commands

What you need

Start it

What happens when you write

Choose the session

What the AI can do when woken

Keep an open session listening

One connector per connection

What it prints

When something does not work

​Commands

​What you need

​Start it

​What happens when you write

​Choose the session

​What the AI can do when woken

​Keep an open session listening

​One connector per connection

​What it prints

​When something does not work

Commands

What you need

Start it

What happens when you write

Choose the session

What the AI can do when woken

Keep an open session listening

One connector per connection

What it prints

When something does not work