Capabilities & Tools

Everything the assistant can do — send email, read the calendar, control your machine, fetch stocks, post to Slack — is a capability exposed as one or more tools. Each tool has a name, a description, and a typed set of parameters. The assistant chooses the right tool for your request and runs it.

How a tool is defined

Conceptually, every tool declares:

Part	Meaning
Name	A stable identifier, e.g. SEARCH_WEB
Description	What it does (used to match it to your request)
Parameters	Typed inputs, some required
Result	A consistent { success, message, data } shape

Browse them all in the Capability Catalog or the per-tool Tool Index.

Picking the right tool

With hundreds of tools available, the assistant never considers all of them at once. It ranks tools by relevance to your request and only weighs the best matches — then selects and runs one.

Relevance ranking is one of the paths that benefits from native acceleration, so selection stays fast even as the catalog grows.

Safety

Capabilities carry a safety flag. Trusted built-in capabilities run without prompting. Anything the assistant creates for itself at runtime defaults to requiring your approval before it runs — you'll see a confirmation first.

Always-available core

A small set of core capabilities is never unloaded, so the assistant can always plan, control the system, and extend itself — even under memory pressure:

Core capability	Role
Self-extension	Creates new capabilities on demand (details)
System tools	Core OS operations (details)
Browser automation	Drives a browser for web tasks

Capability areas

Extending the set

The catalog isn't fixed. The assistant can write new capabilities on the fly when a request needs one, and external tools can be connected via MCP servers.