
# Connect Claude, ChatGPT or Gemini

Your AI assistant can work in Karmaflow.ai as you. You sign in with your own Karmaflow.ai login, choose the workspace, and allow what the assistant may do. From then on it acts with your role's permissions, only while you are a member of the workspace, and everything it changes is recorded as yours, "via" the assistant ("Pat Example via Claude").

There are two addresses to connect to. They use the same sign-in and the same safeguards; they differ in how much of the workspace the assistant can reach.

| | Everyday connection | Workspace connection |
|---|---|---|
| **Address** | `https://mcp.karmaflow.ai/mcp` | `https://mcp.karmaflow.ai/mcp/workspace` |
| **Tools** | About 30: contacts, deals, tasks, notes, meetings, helpdesk tickets, documents, analytics tabs | Every tool your role allows, over 200: the everyday ones plus agents, orchestrations, knowledge, sent emails and texts, calls and chats, approvals and audit, spend, inventory, jobs, websites |
| **Built for** | Day-to-day work with a deliberately small reach | Setting up, configuring, reviewing and auditing the workspace |
| **Where you find it** | The Karmaflow.ai apps listed in assistant directories, or added by hand | Added by hand as a custom connector or custom app |
| **Orientation** | The assistant is told how records and permissions work | The assistant is told to read `get_workspace_guide` first: what this connection may do, which tools do what, and the order to set up an account in |

```mermaid
flowchart TD
    Q{What will the assistant do?} -->|Look things up, log notes, move deals, answer tickets| E[Everyday connection<br/>mcp.karmaflow.ai/mcp]
    Q -->|Build agents and orchestrations, curate knowledge, review and audit| W[Workspace connection<br/>mcp.karmaflow.ai/mcp/workspace]
    E --> L[Your role, the workspace's AI access settings and what you allowed decide what it may do]
    W --> L
```

## Which one fits which job

These are examples. Every business divides the work differently, and the same person may want one connection for some days and the other for others.

**The everyday connection suits** people whose work is customers and records, where a small, predictable set of tools is the point:

- A salesperson who asks "what's due this week?", adds a note after a call and moves a deal to the next stage.
- A support agent who finds a ticket, reads its thread and adds an internal note or a reply.
- A manager who runs a shared analytics tab before a meeting.
- A team where the workspace admin wants assistants to do routine work and nothing that changes how the workspace runs.

**The workspace connection suits** the people who set the workspace up, keep it running and watch over what the AI does:

- An admin setting up a new account: knowledge first, then the voice, chat, SMS, email and task agents, then the orchestrations that connect them.
- An operations lead changing an agent's instructions, checking `get_agent_promotion_status`, and promoting the change when it is right.
- A person above the loop reviewing what the AI did and decided: queued actions waiting for approval, actions it handled on its own (with undo where an action can be undone), and knowledge it proposed.
- An auditor reading audit events, security findings and AI spend.
- Anyone who needs the sent emails and texts, calls and chats, inventory, jobs or website drafts that the everyday connection does not reach.

**Nothing here grants extra rights.** On either address the assistant can do only what three things allow together:

1. **Your role** in the workspace: the same permissions the app's own pages check. A tool your role cannot use is not offered.
2. **The workspace's AI access setting** for external AI agents, per role and per area (None, Read, Read & change), under **Settings (the gear icon) → Agents & AI → AI access** *(sign-in required)*. A workspace admin can, for example, give sales roles **Read** on agents and **Read & change** on CRM.
3. **What you allowed** on the consent page. Allow less, and the assistant asks you to reconnect when it needs more.

A workspace admin's own access is never narrowed by the AI access setting. Their role already allows everything.

## Connect an assistant

You need a Karmaflow.ai login in the workspace, and the workspace's sign-in connections switched on. They are on unless a workspace admin turned them off. You do not need an API key.

**Claude**

1. In Claude, open **Settings → Connectors** and choose **Add custom connector**.
2. Paste the address you chose and add it.
3. When Claude asks, sign in to Karmaflow.ai, choose the workspace, read what Claude will be able to do and press **Allow**.

**ChatGPT**

1. In ChatGPT, open **Settings → Plugins → Browse directory**, then **Add** → **Add custom MCP server**. If **Add** is missing, turn on **Developer mode** under **Settings → Security and login** first.
2. Paste the address you chose as the **Server URL** and leave **Authentication** on **OAuth**: there is no client id or secret to copy.
3. Press **Connect** when ChatGPT offers it, then sign in, choose the workspace and press **Allow**.

**Gemini**

Gemini connects as a *custom app*. Google offers custom apps to personal Google accounts (not work or school accounts), for people 18 or over in the US, in English, with Gemini History on.

1. On a computer, open Gemini on the web, then **Settings → Connected Apps** (or **Personal Intelligence → Connected Apps**). Under **Custom apps**, choose **Add a custom app**.
2. Paste the address you chose and press **Next**. Leave any client id and secret fields empty.
3. Sign in to Karmaflow.ai, choose the workspace and press **Allow**.

Gemini asks you to confirm each action that changes data. To make sure a prompt uses Karmaflow.ai, type `@` and pick it.

### Switching between the two

The address is fixed when you connect: an assistant connected to the everyday address keeps the everyday tools. To change, remove the connection and connect again to the other address.

You keep one connection per assistant and workspace. Where an assistant identifies itself the same way for both addresses, as Claude does, connecting to the second ends the first.

## What stays the same on both

- **It acts as you.** What it creates or changes shows your name with the assistant beside it, in the record's history.
- **Changes to how the workspace runs go through the same steps as in the app.** Agents an assistant creates or edits are drafts until `promote_agent` makes them live. Orchestrations are saved paused, and switched on once they validate. Prompt components go live when pushed. Proposed knowledge is a draft unless your workspace's [AI autonomy](/help/workspace-administration/ai-autonomy) settings let it publish.
- **Personal data rules apply.** The workspace's redaction setting for AI assistant connectors applies to results. Reading the words of one call, chat, email or text is off until a workspace admin turns it on, never available in a workspace marked as healthcare, and card, bank and government-ID numbers are always taken out.
- **You can end it at any time.** Remove it under your user menu → **Account Settings** → **Connected AI assistants** *(sign-in required)*. Workspace admins see every connection, can remove any, and can switch sign-in connections off for the whole workspace, on the **MCP Server** card under **Settings (the gear icon) → Access & security → Tenant API Access** *(sign-in required)*.

## Not available through an assistant yet

Email signatures, inviting users and changing roles, phone numbers and sending domains, and organisation settings (business profile, time zone, business hours, notification routing) are set in the app for now.

## Troubleshooting

**The assistant says to reconnect to allow something.** You did not allow that area when you connected. Remove the connection and connect again, allowing it on the consent page.

**A tool you expected is missing.** Your role does not include it, or the workspace's AI access for that area is **None** for your role (or **Read** for a tool that changes things). Ask a workspace admin. On the everyday address, the tool may simply not be in its short list: use the workspace address.

**The consent page says your role allows none of what the assistant asks for.** The workspace's AI access for external AI agents is **None** in every area your role uses, or sign-in connections are off. A workspace admin can change either.

**Gemini shows no Add a custom app.** Your Google account or location is outside Google's current limits (above).

## Related

- [MCP Servers](/help/integrations/mcp-servers): the per-workspace endpoint with an API key, Claude Code, and connecting your agents to other MCP servers.
- [AI autonomy](/help/workspace-administration/ai-autonomy): what the AI may do without asking.

[Sign in to Karmaflow.ai](https://omni.karmaflow.ai/auth/login)
