# Getting Started 👋

## Installation & First Steps

## Setup API keys

## Chatting with Alice

## Summon Alice

## Agents

## Skills

## Customizing Themes

## Image Generation

## Online Mode

## Vision Skills

## Working with documents & files

## Artifacts

## Code Interpreter

## Multi-agent

## Memory

## Deep Links (x-scheme-url)

# Remote Actions ✨

## Remote Actions Basics

## First Remote Action Scenario

## Quick Scheduling Links Automation

## Using existing automations

# Advanced 📚

## Local Server

## Managing messenger inboxes

## Using Model Context Protocol (MCP)

## Using Zapier MCP

## Using N8N MCP

## Custom Apps

## Open-source models with Ollama

## Connect Alice to OpenRouter.ai

# Inspirations 💡

## Accept Offers with Alice

## Craft Compelling Copy

## Customer Service

## Translations and readability

## Formal Communication

## Generate Images & use Vision

## Getting Mailchimp Campaign Stats

## Google Analytics Reports

## Harnessing Alice for Development

## Instant Text Correction

## Instant Voiceovers

## Instant Zoom Meetings

## Manage Subscribers in Mailer Lite

## Managing Social media

## Managing Tasks

## Quick Google Calendar Events

## Recent News with Perplexity

## Scheduling Social Media Posts

## Simple Automation for Data Insights

## Streamline Project Management

## Translations and Learning

## Unsplash Photos

## Wolfram Alpha

## YouTube Descriptions

## Zoom & Google Calendar

## What is MCP?

The Model Context Protocol is a method for connecting your assistants to "MCP Servers," which can operate locally on your Mac or PC, or on remote servers. These servers provide tools that your assistant can access while chatting with you.

Here we have an example of Spotify MCP. The moment I asked to switch the music, my assistant searched the Spotify catalog and started playback.

## MCP Support

Alice supports the Model Context Protocol almost entirely, including tools, resources, prompts, and sampling features. Tool usage is available for **any model** capable of generating JSON. While JSON Mode and Function Calling capabilities **are not required**, it is still recommended to use state-of-the-art (SOTA) models with MCP features.

To work effectively with MCP Servers, after connecting to them, go to the assistant's settings and limit the number of tools to a minimum. This helps prevent overwhelming the model with context that could degrade its performance and accuracy.

## MCP Configuration

You can connect two types of MCP servers: SSE and STDIO. It is important to note that you need to specify **absolute paths for the executable and server script**, as otherwise Alice will not be able to connect to your MCP Server.

So for example, if we have Firecrawl MCP, we need to first:

1. Install it with `npm install -g firecrawl-mcp` command
2. And obtain the absolute paths for Node and NPX by running the commands `which node` and `which npx` in a terminal. So instead of **npx** you will get something like `/Users/overment/.nvm/versions/node/v22.11.0/bin/npx`
3. Finally, you need your Firecrawl API key

Having all this information, go to `Settings → MCP` and add a new server with the type "STDIO" Then add the Runtime Executable, Server Script path, Arguments, and your API key as an Environment Variable. After that, you may change the server to "Enabled," and if everything is okay, your server may be connected with the given assistant(s).

Another example might be an HTTP SSE server, but in this case, the workflow is much simpler because you only need to ensure that the server is up and running and that you provide the correct URL along with the appropriate HTTP headers.

### How to Use MCP

When you successfully connect the MCP, make sure to assign it to the assistant. Then, when you activate the given assistant, you'll see an indicator in the footer that will inform you about the MCPs that are currently active.

To call any tool or tools, simply tell the assistant what you want to achieve.

**Important Note:** Try to be as precise as possible. This way, you reduce the risk of the LLM using the tools in unintended ways. A good idea is to **limit the available tools to the minimum** and keep the tasks short and simple.

Below is an example of using two MCP Servers—Firecrawl (for accessing web pages) and FileSystem (for accessing local files). I requested access to a certain website and asked to save the pricing information to a Markdown file. As you can see, the assistant correctly identified which tools to use and in what order.

### How it works under the hood?

The logic behind deciding whether to use the tool or not is straightforward. When any MCP Server is active for the given assistant, sending a message triggers an additional prompt to the model. This prompt asks if any tool needs to be used. If so, the assistant enters a loop (with a maximum of 15 steps) in which it continues performing actions until the task is completed or fails.

The important point here is that the currently active assistant's prompt is being included in the "Tool Description" prompt so it will **affect the way the tool is being used**.

Note:

- MCP Servers can be temporarily disabled or enabled with `Command + Shift + P` (you can change this shortcut in General Settings)
- MCP servers are unavailable if Online Mode, Image Generation Mode, or Remote Skill are active