Built-in tools

Tools are how a turn in Cortex Chat reaches past the model: searching the web, reading a page, running a snippet of Python, reading one of your own files, remembering something, drawing a chart from your metrics, or calling an app you connected. Some are built in and always there. Others appear only because of something you set up, and some exist only where the deployment supports them. There is no tools page inside the product to switch these on. A tool shows up as a row inside the answer while it works, and disappears into the result when it finishes. This page covers what is built in, why a tool is sometimes simply absent, what happens to a tool’s result before the model sees it, the Python sandbox in detail, and where the tools you add yourself come from.

​

What is built in

Searching your past chats exists as a capability behind an explicit consent, but that consent is not exposed anywhere in the product yet, so no chat can use it today.

​

When a tool is not offered

A tool is only offered when it can actually run. Web search appears only where a search stack is configured. Python appears only where an executor exists. Business metrics appear only once a metrics source is connected. Where one of those is missing, the tool is not registered at all. You do not get a broken tool or an error row, you simply never see it offered. If you expected a capability and the model says it cannot do it, the honest answer is usually that the tool is not available on your deployment rather than that the model refused.

​

What happens to a tool result

Every result is handled the same way, whichever tool produced it, and in this order. A tool call ends in one of five ways: it succeeded, it errored, it timed out, it was refused, or it was interrupted. A row that produced nothing at all reads No output.

​

The Python sandbox

Chat’s Python tool is a snippet sandbox, not a build machine. It has 256 MiB of memory, no network, and a few minutes of wall clock. It is the right tool for arithmetic, a data frame, a plot or checking a piece of code, and the wrong one for installing packages or building a project. Two practical rules:

• Print what you want to see. The last value is discarded, so a bare expression on the final line gets you nothing.
• Files the run writes to its outputs/ folder are collected into your Library when the run finishes. Direct files only, at most 10 files and 10 MiB in total. A run that failed collects nothing.

If you need a shell, a repository and a real workspace, that is Cortex Code, not a chat.

​

Skills are loaded, not always on

Skills are recipes the model can pull in. It does not carry them all at once: it sees a one-line index of the skills available to it, and fetches the full body of one only when it decides to use it. That is why a skill can change how an answer is written without you naming it in the prompt. Typing / in the composer lists the skills turned on for Chat.

​

Tools you add yourself

Everything above is built in. Three more sources of tools are yours to set up, and each has its own page.

​

Related

• How Chat works - how a turn is assembled and how tools enter it.
• Library - where stored results and sandbox outputs end up.
• Deep Research - a research run, which replaces the ordinary tool loop.
• Scheduled tasks - runs that get no tools at all.
• Models and thinking - the chip where thinking and Deep Research are switched.