---
id: con-tools-and-connectors
schema: concept/1
layer: garden
status: established
name: Tools and connectors
aka: ["connectors", "integrations", "apps in ChatGPT", "extensions", "plugins", "tool use", "actions"]
summary: >
  Ways for an AI to do more than write: search the web, read your email or calendar, look in shared
  drives, or take actions in other apps, using access you grant.
first_seen: "2023-03"
relations:
  builds_on: [con-large-language-model]
  evolved_from: []
  related: [con-privacy-and-data]
  parallels:
    - idea: "giving a temp a building badge and system logins"
      field: "office administration"
      note: "What someone can get done depends on what they've been given access to, and access should match the job."
implementations:
  - { provider: anthropic, name: "Claude connectors" }
  - { provider: openai, name: "ChatGPT apps and connectors" }
  - { provider: google, name: "Gemini apps (formerly Extensions)" }
  - { provider: perplexity, name: "Perplexity connectors" }
competencies: [cmp-describe-the-task, cmp-use-responsibly]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "The step from AI that talks to AI that reaches into your real work; prerequisite for agents and automation."
---

On its own, a language model can only work with what is in the conversation. Tools let it reach further. A search tool lets it look things up. A calendar connector lets it see when you're free. A drive connector lets it read the documents you point to. Some connectors can also act: draft an email, create a calendar event, update a spreadsheet.

You decide which tools are switched on, and good products ask before anything is sent or changed. A sensible rule is least access: connect what the task needs, read-only where possible, and disconnect what you no longer use.

Connectors also change what "checking the result" means. When the AI cites your own files, you can open them and confirm. When it acts in an app, you open the app and look at what happened.

## Example

A project coordinator connects her calendar and asks, "Find three 45-minute slots next week when I'm free in the afternoon, avoiding Fridays." The AI reads her calendar and proposes times. She chooses one and creates the invite herself.

## How it connects

- **MCP** is a shared standard that makes building connectors easier across tools.
- Tools are what let **agents** and **automation** take real actions.
- Every connection is a **privacy and data** decision about what the AI can reach.

## Try it

Open your AI tool's connectors or apps settings. Before switching anything on, write down what you'd want it to read and what it should never touch. Connect one read-only source and ask a question you can verify.