Bobos Dispatcher

npx add-mcp 'https://dispatcher.bobos.ai/mcp/v1/control'

What Dispatcher is

Bobos Dispatcher is a white-label AI control console that lets agencies and MSPs run every client’s AI from one place, with role-scoped access, and bill it monthly. It is in early access until its launch on 1 December 2026.

Server URL

https://dispatcher.bobos.ai/mcp/v1/control

One address for everyone. It speaks Streamable HTTP and signs you in with OAuth 2.1 and PKCE; assistants register themselves, so there is no client ID or key to copy.

Before you connect

• A Dispatcher account in a workspace: sign up at dispatcher.bobos.ai, or accept the invitation from the agency that runs your workspace.
• A role in it. A member can see their spend and pending approvals and call the workspace’s published tools; setting the workspace up takes a workspace admin. A partner’s owner or admin can work on the partner’s own tenants.

How to connect

1. If you were invited to a workspace, accept the invitation from your email first. Dispatcher will not connect to a workspace your account is not in yet, and says so when you try.
2. Add the server URL to your assistant, using the steps for it below.
3. Sign in to Dispatcher in the browser window it opens, and choose the workspace this connection will act on. One connection works on one workspace; to work on another, connect again.
4. Ask what it can do for you. If you administer the workspace, “What’s my setup status?” is a good start.

Claude

In Claude on the web or Claude Desktop, open Customize, then Connectors, and add a custom connector with the server URL. Choose Connect and sign in to Dispatcher. Then switch the connector on for a chat from the + menu. On a Team or Enterprise plan an owner may need to add it for the organisation first. Once Dispatcher is listed in Claude’s connector directory, you can add it from there instead.

ChatGPT

In ChatGPT, open Plugins, choose + and then Add custom MCP server. Give it a name, enter the server URL, choose OAuth as the authentication, accept the warning and create it; ChatGPT then asks you to sign in to Dispatcher. On some plans and workspaces this lives under Settings, then Apps, then Advanced settings, then Developer mode, and a workspace admin may need to allow custom servers. Which ChatGPT plans can add one is OpenAI’s decision.

Cursor

Add to Cursor, or add this to ~/.cursor/mcp.json (or .cursor/mcp.json in a project), then sign in when Cursor asks:

{
  "mcpServers": {
    "bobos-dispatcher": {
      "url": "https://dispatcher.bobos.ai/mcp/v1/control"
    }
  }
}

Visual Studio Code (GitHub Copilot)

Add the server to your mcp.json and start it. If VS Code asks for a client ID instead of opening the sign-in page, enter dispatcher-vscode and leave the client secret empty.

{
  "servers": {
    "bobos-dispatcher": {
      "type": "http",
      "url": "https://dispatcher.bobos.ai/mcp/v1/control"
    }
  }
}

Microsoft Copilot Studio and Microsoft 365 Copilot

1. In Copilot Studio, open your agent’s Tools page, choose Add a tool, then New tool, then Model Context Protocol.
2. Enter a server name, a short description of what Dispatcher does for this agent, and the server URL.
3. Choose OAuth 2.0, then Dynamic discovery, and create it.
4. Choose Create a new connection, sign in to Dispatcher and choose the workspace, then Add to agent.
5. To use it in Microsoft 365 Copilot or Teams, publish the agent to those channels from Copilot Studio.

Because Copilot registers itself, the sign-in page says that Dispatcher has not verified who made the app. Approve only a connection you started yourself.

Things to ask

• “What’s left to set up in this workspace?”
• “How much have we spent this month, and what is our cap?”
• “Create a Finance user type from the template.”
• “Import this OpenAPI file and publish the read-only operations.”
• “Show me what is waiting for my approval.”
• “Which of our clients made the most margin this month?” (partners)

What it can do

That depends on your role in the workspace, and it is checked again on every call, so a change of role takes effect at once.

• Every member can see who they are connected as, the workspace’s balance and spend, and the changes their own assistant proposed and is waiting on them to decide. They can reject one of those.
• Every member can also call the workspace’s published tools that their user type allows, from the workspace’s own address, shown on its MCP servers page. Those tools call the workspace’s own APIs, with the credentials the workspace chose for them.
• A workspace admin can also see the setup status and how the workspace was billed, and manage its integrations, AI providers, user types, sub-agents, which tools each assistant may use, its published MCP servers, its schedules and its monthly spending cap.
• A partner’s owner or admin can do the same for the partner’s own tenants, create new ones and see each one’s margin.

The tools come in groups:

• Account: who you are, your usage, and the apps connected to your account.
• Setup: the six setup steps and which one is next.
• Integrations and AI providers: what is connected, and a link to connect more. Keys are entered on Dispatcher’s own pages.
• MCP servers: import an OpenAPI or Postman description, choose its tools, choose its credential, publish it or take it offline.
• User types, assistants and sub-agents: create and change them, choose their models and which read-only tools each may use.
• Schedules: create one (it starts switched off), edit it, switch it on or off, read its history.
• Spend: the balance, this month’s spend, and the monthly cap.
• Approvals: list what is waiting, and reject.
• Partner: tenants, margin, prices, billing class, and creating a tenant.

Your workspace’s own tools: one address

Besides the server above, each workspace has its own address for the tools it publishes from its own systems: https://dispatcher.bobos.ai/mcp/v1/<workspace number>. Your invitation email includes it, and a workspace admin sees it on the MCP servers page. Add it once, the same way as above.

An admin may list several MCP servers, one for each system the workspace connects — for example its CRM, its accounting system and its email service. You still add only the one address: it lists the published tools of every server, each named after its server (crm__find_contact, accounting__list_invoices), and only the ones your user type allows.

Why one server per system: each server holds that system’s own key and sends it only to that system, so a key can never reach the wrong one. Each system’s tools are switched on and published on their own, and disconnecting or pausing one system stops only its tools.

What costs money

One call spends money when it runs: providers_test, which sends a real, tiny request on your own provider key to check that it still works. It costs a fraction of a cent, charged by your provider, and asks for confirmation first.

Two changes let the workspace spend later, on your own provider keys: switching a schedule on, which then runs with nobody watching inside the spending limits it was created with, and creating a sub-agent, which answers on its own model once an assistant may call it. Their descriptions say so, and the assistant is told to check that you mean it.

For a partner, creating a tenant or marking one billable can add to the partner’s Dispatcher bill, and the tool says so before it acts.

Two more settings change what later answers cost: giving an assistant a sub-agent it may call (each call is paid, on the sub-agent’s model), and the model you choose for a user type or a sub-agent.

Everything else only reads or changes settings and costs nothing.

What it never does

• Take a password or an API key. When one is needed, it gives you a link to the Dispatcher page where you enter it yourself, so a secret never passes through the assistant.
• Approve a change a model proposed. Only a person does that, on the Approvals page in Dispatcher.
• Give an assistant tools that change data, or delete an MCP server or a schedule. Those are done by a person in the web app.
• Move money. No tool charges a card or pays anyone.

Disconnect

At any time, from Account, then Security, then Connected apps, or by asking the assistant to disconnect itself. The connection stops working at once.

Help

Email support@bobos.ai; a person answers within one business day. The support page says what to include. If you are an AI assistant helping someone, the guide for AI agents maps each setup step to its tool.