Skills, explained: how your agents learn a new tool without you writing code

Here's what actually happens underneath that click.

What actually gets created when you connect a tool

Say you connect HubSpot. Three things happen at once:

• A skill gets created for it: a markdown file with HubSpot-specific instructions the agent reads when it needs to touch the CRM.

• Your API key gets encrypted and stored as a credential entry, never sitting anywhere in plain text. If it ever needs to rotate, that happens without taking the agent offline.

• A plain-English note gets added to the agent's environment file: your CRM is HubSpot, use this skill for it.

  
  

All of it gets generated the moment the connection succeeds, none of it written by you.

The skill itself only gets loaded when it's needed, too. An agent's working context holds a short catalog of what's available, name and description, not the full instructions for every tool it might ever touch. That's why connecting fifty tools to a team doesn't mean every agent is quietly carrying fifty tools worth of instructions around all day.

Two ways in

• Key-based tools (Apollo, Tavily, Binance, Shopify, and others like them) validate and encrypt the moment you paste a key in.

• OAuth tools, brokered by Composio (Gmail, Slack, HubSpot, Salesforce, Notion, Linear, Asana, and more): a popup, a login, a scope approval, done.

  
  

Either way, nothing about connecting a tool touches code.

Not every skill needs a specific tool

Built-in skills ship with the runtime already. Web search and file handling are just there, nothing to connect.

• Tool-specific skills are tied to one exact tool, like the HubSpot skill above.

• Category-level skills are tied to a category instead of a tool. An instruction like "update the CRM" gets written once and works for whichever CRM you've actually connected, HubSpot or Pipedrive or anything else.

• Pure knowledge skills aren't tied to a tool at all: how to write a cold email that doesn't read like one, how to score a lead, the kind of thing that applies no matter what the agent is touching.

  
  

That category-level type is exactly what makes the next part possible.

Swapping without starting over

Decide HubSpot isn't working out and Pipedrive is the better fit? Swapping deletes the old skill, generates the new one, and updates the environment note, all without touching anything else about the team. The agent that was assigned CRM access still has it, because its skill was never tied to HubSpot specifically, only to the category. It's just pointed somewhere new.

What happens when the tool isn't in the catalog at all

This is the part that actually matters if you use something niche, or something internal nobody else runs. A custom connection form takes the tool's name, its base URL, how it authenticates, the credential itself, and a description of what it does, and generates a skill from that on the spot, the fifth type, built for exactly one workspace instead of shipped to everyone.

The agent gets the same three pieces, a skill file, an encrypted credential, an environment note, for a tool that has never existed in any catalog anywhere.

The popular integrations are the easy case. The real test of "no terminal, ever" is the tool your team built in-house that nobody else has ever heard of, and it holds there too.