Use Aito from Claude Code and other AI assistants

An AI coding assistant can call Aito directly. aito-mcp is an MCP server that gives the assistant Aito's query operators as tools, and the Claude Code plugin adds a skill that walks the assistant through adding a predictive feature to your application. Both are built to say when Aito fits and when another tool is the better choice, following When to use Aito, and when not.

1. Get the server

With uv, there is nothing to install: uvx aito-mcp fetches and runs the aito-mcp package from PyPI. Or install it with pip, next to the Aito Python SDK:

pip install 'aitoai[mcp]'

Either way it needs Python 3.10 or newer. The server is also listed in the MCP registry as io.github.AitoDotAI/aito.

2. Point it at an instance

aito-mcp finds your instance the same way the Python SDK's aito.Client() does:

• AITO_URL and AITO_API_KEY in the environment, or
• the profile that aito start writes when you run Aito locally.

A read-only key is enough for everything except writes.

3. Connect your assistant

Claude Code:

claude mcp add aito -- uvx aito-mcp

(or claude mcp add aito -- aito-mcp if you installed it with pip).

Any other MCP client runs aito-mcp as a stdio server. In a client that reads a JSON server list, the entry looks like this:

{
  "mcpServers": {
    "aito": { "command": "uvx", "args": ["aito-mcp"] }
  }
}

What the assistant gets

The tools are predict, recommend, relate, match, search, query, evaluate, get_schema, put_schema and upload_rows. Each one takes a request body written the way the v2 API reference writes it, and each tool's description says when to use it and when not.

• Evidence on every answer. predict, recommend and match return the value, its probability $p and $why, the factors that moved it, strongest first.
• Thresholds from your own data. evaluate scores a query on held-out rows, so the assistant can choose where to act, where to ask a person and where to step back.
• Engine errors pass through with the engine's own message, so the assistant can fix the request.

Read-only by default

The server refuses writes (put_schema, upload_rows) unless you start it with writes allowed:

AITO_MCP_ALLOW_WRITES=1 aito-mcp

Keep the default when an assistant works against shared or production data. Allow writes for a local or sandbox instance where you want it to create the schema and load rows.

The Claude Code plugin

The plugin bundles the server with a skill, add-predictive-feature. The skill walks the assistant through:

1. a fit check that stops and names the better tool when Aito does not fit;
2. how to keep your customers' data apart, if your application serves several;
3. a schema derived from your existing table;
4. a first _predict;
5. a threshold chosen with _evaluate on your own held-out rows;
6. act, assist and abstain wiring in your code.

Install it in Claude Code from the plugin marketplace in the aito-python-tools repository:

/plugin marketplace add AitoDotAI/aito-python-tools
/plugin install aito@aito

Next

• When to use Aito, and when not: the fit and non-fit cases the tools and the skill follow.
• Benchmarks: the evidence behind each case, including where Aito loses.
• v2 API reference: the request bodies the tools take.

Back to developer docs