DialogBrain

npx add-mcp 'https://api.dialogbrain.com/mcp/'

Using DialogBrain from Claude, ChatGPT and Cursor

Your workspace is an MCP server. Point an assistant at one URL and its inbox, contacts, calls, tasks and knowledge become tools that assistant can call — the same tools your own agents use, with the same permissions.

Updated 2026-09-29

§ Before you start

• A DialogBrain workspace. The connection is scoped to one workspace, and everything the assistant reads or sends belongs to it.
• The client you want to connect: Claude, ChatGPT, Cursor, VS Code, Devin Desktop (formerly Windsurf), Claude Desktop, Claude Code, Codex, Gemini CLI, Antigravity or Grok. Any client that speaks MCP over HTTP works, including one you wrote yourself.
• Nothing to install. There is no package and no local process: the server is ours and it is reached over HTTPS.

1. Copy the server URL

   Open Settings → Developer. The Connect section holds one address, and it is the same for every client and every workspace:

   https://api.dialogbrain.com/mcp/
   

   Which workspace you land in is decided by who authenticates, not by the URL. That is why the address is not a secret and the credential is.
2. Step 2

   Add the server to your client

   Settings → Developer. Every client the guide covers has its own entry, and each one prints the snippet to copy. In Claude on the web: Settings → Connectors → Add custom connector, and paste the URL. In ChatGPT on the web: Plugins → Add → Add custom MCP server, and paste the same URL; the free plan works too. Both then walk you through authentication. In Grok on the web: Plugins → Connectors → New Connector → Custom, give it a name and paste the same URL. Grok opens a popup to sign you in, then lists our tools beside its own connectors. In Claude Code, install the plugin: it adds the server and six skills that know how to set up a widget, connect a channel, build an agent, run the task board and write a lead report. Then run /mcp and sign in.

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

   If you only want the server without the skills:

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

   In Codex CLI and Gemini CLI it is the same shape, and both open a browser to sign you in when you first use the server:

   codex mcp add dialogbrain --url https://api.dialogbrain.com/mcp/
   

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

   Antigravity has no add command and its OAuth for remote MCP does not hand the token on, so it is the one client that takes an API key from the start. Put this in.agents/mcp_config.json in the project (or ~/.gemini/config/mcp_config.json for every project); note serverUrl, not url:

   {
     "mcpServers": {
       "dialogbrain": {
         "serverUrl": "https://api.dialogbrain.com/mcp/",
         "headers": { "X-API-Key": "<your API key>" }
       }
     }
   }
   

   In Cursor, put this in ~/.cursor/mcp.json and restart:

   {
     "mcpServers": {
       "dialogbrain": { "url": "https://api.dialogbrain.com/mcp/" }
     }
   }
   

   In VS Code the file is.vscode/mcp.json in your project and the shape is slightly different — servers rather than mcpServers, and the transport is named:

   {
     "servers": {
       "dialogbrain": { "type": "http", "url": "https://api.dialogbrain.com/mcp/" }
     }
   }
   

   Claude Desktop (v0.9 and later) uses the Cursor shape in claude_desktop_config.json; Devin Desktop — the renamed Windsurf — uses it in ~/.codeium/windsurf/mcp_config.json, or ~/.config/devin/mcp_config.json on newer builds. Settings → Developer prints the exact snippet for each, so you can copy rather than retype.
3. Step 3

   Authenticate

   Prefer OAuth where the client offers it: add the server, and the client opens a browser to sign you in. In Claude Code that means running /mcp after adding the server. Nothing is stored in a config file, and access follows your account. Where OAuth is unavailable (Antigravity, and any client whose sign-in never completes), use an API key. Settings → Developer → API Keys creates one; it is shown once, so copy it then. The key goes in an X-API-Key header:

   {
     "mcpServers": {
       "dialogbrain": {
         "url": "https://api.dialogbrain.com/mcp/",
         "headers": { "X-API-Key": "your-api-key" }
       }
     }
   }
   

   A few clients accept only a bare URL with no headers. For those the same key can ride in the query string as?token=… — the Any MCP client (HTTP) entry in Settings → Developer prints that form. Cursor is the one to watch: versions that skip the OAuth prompt answer "unauthorized" and give no other clue. If that happens, switch that entry to the API-key variant above.
4. Step 4

   Check that it took

   Ask the assistant what it can do with DialogBrain, or to list its DialogBrain tools. A connected client reports a large tool surface, grouped by what it acts on: messages, threads, contacts, calls, agents, tasks, knowledge, files, analytics, widgets. Then ask for something small and read-only — "find the contact called …", "what were the last messages in that thread" — before you let it write anything. A tool that reads proves the connection and the workspace in one step. One thing that looks like a failure and is not: the Claude Code Channel entry in the same settings page is a channel, not a tool server. Other MCP clients connect to it and correctly find no tools. If a client reports an empty tool list, check which of the two you pasted.

What you can do once it is connected

Read and answer. The assistant can search threads and messages, open a conversation, and send a reply through whichever channel that conversation belongs to — Telegram, WhatsApp, Instagram, email — without you naming the channel.

Work the contact base. Find people, merge duplicates, update fields, look at a contact profile with its history across channels.

Drive agents and calls. Start a call, read its transcript afterwards, hand a conversation to an agent or take it back, create and assign tasks on the workspace board.

Ask questions of your own data. The knowledge base and the analytics surface are tools too, so "how many conversations did we answer last week" is a question the assistant can actually answer rather than guess at.

Two things the server tells your client, and why they matter

Every response carries the workspace it ran in. A reconnect starts a new session in the API key’s default workspace, so a session can quietly move: the first symptom is a tool saying an id you just used does not exist. The fix is to switch workspace, not to retry the call or doubt the id.

Every response also carries a version of the tool surface. We add and change tools on deploy; when that value changes mid-session, a client holding cached schemas should re-read the tool list rather than trust the parameters it memorised at connect.

Both are sent to the client automatically at connect time. You do not have to do anything with them, but they explain two failures that otherwise look like bugs in your prompt.

§ Q & A

Is the server URL a secret?

No. It is the same address for everyone, and it is printed in the settings page. The credential is the secret: whoever authenticates decides which workspace the session sees.

OAuth or an API key?

OAuth wherever the client supports it. Nothing lands in a config file, access follows your account, and revoking it is a normal account action. Use a key when the client has no OAuth flow, or when the connection belongs to a machine rather than a person.

My client connected but found no tools.

Almost always the Claude Code Channel URL pasted where the MCP server URL belongs. That endpoint is a transport for Claude Code sessions and deliberately registers no tools. The server address is the one under Connect.

Cursor says "unauthorized" and nothing else.

Some Cursor versions never show the OAuth prompt and fail silently. Use the API-key form with the X-API-Key header instead; it works on every version.

Can several people connect at once?

Yes. Each person authenticates as themselves, so what they do is attributed to them, and sessions are independent — one person switching workspace does not move anybody else. Where a surface has role-based limits, such as the task board or workspace settings, they apply to whoever is calling.

What happens when I revoke a key?

Every integration using it stops immediately, and there is no undo. Give machines their own keys so revoking one does not take down the rest.

Can I connect my own MCP servers the other way round?

Yes, and that is a separate feature: an agent can be given any MCP server or HTTP API you connect, and those tools sit beside ours at the same rank. This guide is about the opposite direction — your assistant reaching into DialogBrain.