---
title: "MCP overview"
description: "Connect Claude, Cursor, ChatGPT or any MCP client to your Lookio workspace so an AI agent can import resources, build assistants and ask questions."
canonical: "https://lookio.app/docs/mcp/overview"
updated: "2026-11-03"
---

# MCP overview

Lookio runs a Model Context Protocol (MCP) server. An AI client connected to it can do what the REST API does: import resources, create assistants, ask them questions and, for an admin, manage tags and members. The tools use the same service layer as the app, so a permission refused in the app is refused to the agent.

```text
https://mcp.lookio.app/mcp
```

## Two servers

| Server        | URL                               | Tools                                                                                                                                                      |
| ------------- | --------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Standard**  | `https://mcp.lookio.app/mcp`      | One task-shaped tool per action: `list_resources`, `add_site`, `ask_assistant`. The right default for most clients. See [Tools](/docs/mcp/standard-tools). |
| **Code Mode** | `https://mcp.lookio.app/mcp/code` | Two tools, `search` and `execute`, for agents that prefer to write a small script. See [Code Mode](/docs/mcp/code-mode).                                   |

Both authenticate the same way, enforce the same scopes and share the same rate limit.

## What an agent can do

- Add a text, a web page or a whole site, check that it was read, and read back what was indexed.
- Create an assistant, choose what it reads and change its context.
- Ask an assistant a question and get the answer with its citations, continue it with its `conversation_id`, and read, rename, delete or rate your conversations.
- Put an assistant on a website with a [widget](/docs/widget/overview), and read what visitors and your product's users asked in the [logs](/docs/platform/logs).
- Read the account, the credit balance and the usage.

Asking is the only tool that spends credits. Every tool says what it costs in its description, so an agent knows before it calls.

## How you connect

1. Add the server URL to your client. See [Connect an AI client](/docs/mcp/connect).
2. The client opens Lookio in your browser. You choose the workspace and approve the permissions.
3. Lookio creates a personal connection that you can review and revoke in **Developer tools**.

You can also give a client a workspace [API key](/docs/platform/api-keys) in an `Authorization: Bearer` header, which is useful for unattended agents.

## Permissions

An agent never gets more than the person who connected it, and never more than the scopes you approved. Tools it may not use are not even listed. See [Safety and permissions](/docs/mcp/safety-and-permissions).

## Transport

The server speaks MCP over Streamable HTTP and holds no session between requests. It answers clients of the 2026-07-28 revision and earlier clients that use stateless Streamable HTTP. The older HTTP with server-sent events transport is not supported. For a client that can only start a local process, the CLI ships a bridge: `lookio mcp`.