Dialogbrain

DialogBrain for Claude Code

Connect WhatsApp, Telegram, email and a chat on your website to one inbox, and build AI agents that answer your customers there — all from Claude Code.

DialogBrain is a customer messaging platform: every channel lands in one shared inbox, with contacts, calls, a task board and AI agents on top. This plugin puts that workspace inside Claude Code as tools, plus eight skills that know the order of each setup job.

Install

claude plugin marketplace add saloprj/dialogbrain-claude-plugin claude plugin install dialogbrain@dialogbrain

Then run /mcp inside Claude Code and sign in to DialogBrain in the browser. There is no API key to copy, and the plugin can do exactly what your own account can do, in the workspace you are signed in to.

Things to ask for

Connect my WhatsApp and my work email.

Put a chat widget on example.com. Only allow my domain and keep replies as drafts I approve.

Create an agent that answers customers on WhatsApp and the website chat. Never promise prices. Test it before it answers anyone.

Where did we lose leads this week?

What the skills do

Ask in plain words; the matching skill loads on its own, and only then.

• connect-channel — "connect my WhatsApp / my email / a Telegram bot": connects what needs no sign-in (an IMAP mailbox, a bot token) straight from the chat, and for the rest sends you a link that opens the exact screen.
• setup-livechat-widget — "put a chat widget on my site": creates the widget, restricts it to your domains, hands you the two script tags, and checks that an agent answers.
• create-agent — "an agent that answers my customers": writes it from a description, wires a trigger, sends a test message and reads the trace.
• voice-agent — "a phone agent": creates it, sets the voice, and calls you to test it.
• tasks-board — the workspace task board: assign work to people or agents, track what is overdue, comment with screenshots.
• lead-report — "where did we lose leads this week": queries analytics, reads the threads, and writes one summary.
• first-message — "write to this new number / handle / address": the exact form for a first message on each channel (WhatsApp or SMS to a number, a Telegram @username, email, LinkedIn, Discord, X) and the account it goes from.
• react-to-events — "tell me when this person writes", "watch this group", "notify me on every new lead": narrows the trigger to the messages you care about, then picks what happens — a reply, a note to you, a human takeover, a task on the board.

Letting DialogBrain reach you (not available to everyone yet)

The plugin also ships a channel server, so a DialogBrain agent can hand work to this session instead of answering by itself — a new lead worth judging, a task assigned from a phone.

Read this before trying. Inbound push into a Claude Code session is an experimental feature of Claude Code, and two of its gates are not ours to open:

• the Channels feature has to be enabled for your account;
• the channel has to be allowed. Claude Code allows a channel from its own allowlist, or from your organisation's managed settings (allowedChannelPlugins). A plugin on neither list is refused unless the session is started with --dangerously-load-development-channels, which is meant for local development.

So installing the plugin does NOT give you this. It gives you the tools, the skills and the agents, which need no flags at all. If the steps below end with Claude Code saying the channel "is not on the approved channels allowlist", that is this gate, not a mistake on your side, and nothing in the plugin can lift it.

Nothing happens until you set it up either: with no API key the channel stays off, the session is unaffected, and the plugin's tools work as before.

It needs Node.js 20 or newer on PATH — Claude Code installed as a standalone binary does not bring one. Without it the channel never starts, and the only evidence is a spawn error in the session's own MCP log, so check node --version first.

1. Give the plugin an API key (cabinet → Settings → Developer) and a name for this machine. Claude asks for both when you install the plugin, and the key is stored the way your client stores a secret — nothing is written into this repository or into a shell profile.

If you install from a checkout instead, the same two values can go in ~/.claude/channels/dialogbrain/.env, which the channel reads when the plugin's own settings are empty (for the name, a DIALOGBRAIN_CHANNEL_NAME set in the shell that starts claude wins over both; for the key, a filled plugin key wins over a shell export):

   DIALOGBRAIN_TOKEN=<your key>

DIALOGBRAIN_CHANNEL_NAME=

1. Allow the channel once on the machine, then start Claude Code with pushes enabled (inbound push is off by default, and a running session cannot be switched on). An admin adds this to /etc/claude-code/managed-settings.json:

   { "channelsEnabled": true,

"allowedChannelPlugins": [{ "plugin": "dialogbrain", "marketplace": "dialogbrain" }] }

and every session on that machine starts with the short form:

   claude --channels plugin:dialogbrain@dialogbrain

the same session again, keeping its id:

claude --channels plugin:dialogbrain@dialogbrain --resume

Without the managed settings Claude Code refuses the short form, with the message "is not on the approved channels allowlist". For a quick local try without them, the development flag takes the same plugin entry: claude --dangerously-load-development-channels plugin:dialogbrain@dialogbrain.

The session then appears by itself in DialogBrain under AI Agents → "Claude Code sessions" (and disappears when you close it). Create an agent for it there ("Create agent for this session") or pick it in an existing Claude Code agent's Session card; that agent's triggers send their messages to this session, and its Activity tab shows every call and answer.

1. Ask for it in words: "wake me here when a new lead arrives". The react-to-events skill sets the rest up.

Name the channel in ONE flag only. Listing it in both --channels and --dangerously-load-development-channels registers it twice — once without the development mark — and the check trips over the unmarked copy and refuses the lot.

Use ONE DialogBrain channel per session: if a project's .mcp.json also has a hand-configured dialogbrain-channel, remove it. Two channels register separately, possibly under two different keys, and work can go to the one Claude Code is not listening to.

Only messages from the other side of a conversation trigger a push. What the connected account sends itself is outgoing and wakes nothing.

If nothing arrives, ask DialogBrain which sessions it can see. Being in that list proves only that the channel server reached the server: it opens its socket whether or not Claude Code is willing to deliver anything. The session's own MCP log is the honest answer — it says either "Channel notifications registered" or the reason it skipped them.

What the channel sends us, and when

The channel server is the only part of this plugin that talks to anything on its own, and it stays silent until you give it an API key: with no key it prints one line and connects nothing.

With a key, it opens a WebSocket to the DialogBrain server your key belongs to — https://api.dialogbrain.com unless you point it elsewhere — and introduces the session so a person can address work to this machine by name. That introduction, repeated on every reconnect, is:

• the name you chose for this machine, or hostname:project-folder if you chose none
• the Claude Code session id, the project directory, the working directory, the hostname and the process id
• which Claude Code entrypoint started it and which subagent model it uses
• the current git branch of the project directory
• a one-line title for the session: its summary, or the first thing you typed, cut to 120 characters

That last line is what makes a desktop recognisable in a list of several, and it is the only session content that leaves your machine. Nothing else is read or sent: not your files, not the conversation, not the output. Everything after the introduction travels the other way — the messages your customers send you, pushed into the session.

The key itself is stored by Claude Code when you install from the directory, or in ~/.claude/channels/dialogbrain/.env when you install from a checkout, and is sent only as a header to the server it belongs to.

Channels it can connect

WhatsApp, Telegram, a Telegram bot, Instagram, Facebook Messenger, email over IMAP or Google, a chat widget on your website, and phone numbers for voice calls.

What it will not do

Connecting WhatsApp means scanning a QR code with the phone that owns the number, the way WhatsApp Web works. No assistant should do that for you, so this one does not try: it sends you a link that opens DialogBrain on that step. For a mailbox it asks for an app password, never your main password, and you can revoke it at any time. An agent's replies wait as drafts until you decide it can answer on its own.

Just the server, no skills

claude mcp add --transport http dialogbrain https://api.dialogbrain.com/mcp/

The same address works in ChatGPT, Claude, Cursor, Codex, Gemini CLI and any other MCP client, with OAuth sign-in: https://dialogbrain.com/help/set-up-from-ai-assistant

Requirements

A DialogBrain account (dialogbrain.com) and Claude Code. New workspaces get a free trial, so you can connect a channel and test an agent before paying.

MIT licensed. This repository is published from the DialogBrain monorepo on every release; changes are made there, not here.