# Connect external AIs over MCP

> Reach your permitted agents, published app tools and dashboard from an external AI client.

## Before you begin

Use an active business workspace membership and an MCP client supporting remote Streamable HTTP and OAuth. Claude, ChatGPT, Cursor and VS Code have setup instructions on Botchi's **MCP** page. Availability of custom connectors depends on the external client's plan.

## Connect your client

1. Open **MCP** in the dashboard.
2. Select your AI client to see its setup instructions.
3. Copy the **MCP endpoint** URL or configuration snippet.
4. Add the remote server in your client's MCP configuration.
5. Complete OAuth with your own Botchi workspace account.
6. Ask the client to list the agents and tools you can use.

The endpoint is your Botchi backend URL followed by /api/mcp. Copy the displayed URL rather than constructing one. There are no shared API keys; a client without OAuth cannot connect.

The client signs in as you. Botchi checks your live membership and permissions; removing your membership ends access.

## Check what you can reach

1. Check the listed agents before sending work.
2. Use only published app tools available through your app access.
3. Review your dashboard role before granting an external client operational access.

Every member reaches permitted agents and published app tools. Dashboard access adds tools progressively: viewers read, editors make operational changes, and owners also govern the account.

Dashboard tools include Workspace and agent Homes. Editors can write, delete and share within their permissions.

Editors can also generate images with a saved visual preset, as in [Studio](/en/docs/studio). Images save to the Workspace, the client receives a preview, and each image uses Sparks.

Agent tools let you list agents, send tasks, poll long runs, and approve or reject gated actions. They are named `list_agents`, `send_message_to_agent`, `get_agent_run` and `approve_agent_action`. `search_chat_messages` finds past messages in your personal chat or your chat with one agent; it runs nothing and uses no Sparks.

## Review actions before execution

1. Read the proposed change when a dashboard tool returns `CONFIRMATION_REQUIRED`.
2. Explicitly approve the exact change you want.
3. Let the client repeat that tool call with `confirm=true`.

> **Watch out** — Dashboard confirmation is a contract with the client. Botchi cannot verify that a human approved; treat this connection as access to operate within your role.

Agent approval gates remain separate. An Ask tool pauses the run and sends a push notification. Approve or reject it in Botchi or through the external client. External access does not automatically approve gated actions.

## Monitor cost and limits

1. Check the agent's Activity and [usage reporting](/en/docs/usage) for runs marked MCP.
2. Check the signed-in member's [spend limit](/en/docs/billing/spend-limits) before sending more work.
3. Wait for the indicated retry time if the per-member rate limit is reached.

Runs consume workspace Sparks and are attributed to the signed-in member. An inactive subscription or an empty Sparks pool prevents new work.

| Per-member MCP budget | Per minute |
| --- | --- |
| Calls | 60 |
| Executions | 10 |

The budget covers the member's MCP connection across clients.

## Related

- [Agent tool approvals](/en/docs/agents/tools)
- [App access](/en/docs/apps/access)
- [Dashboard roles](/en/docs/team/roles)