> ## Documentation Index
> Fetch the complete documentation index at: https://docs.qontext.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Context agents

> Agents that create, edit, and synthesize context in your repository

## Overview

Context agents create and edit files in your context repository from a chat prompt. Open the chat on the **Context** tab, describe what you want, and the agent does the work: it builds folder structures, writes and edits files, and moves or removes what's outdated.

Context agents also have access to your synced data. That means you can prompt them to replicate, aggregate, or synthesize context from your connected [sources](/sources): for example, turning scattered source documents into one curated file.

Every edit a context agent makes is a regular change to the repository: versioned, visible in the **Changes** tab, and held for review on files that require it.

***

## Opening the chat

<Steps>
  <Step title="Go to the Context tab">
    Context agents work inside the context repository, so the chat is available on the **Context** tab only.
  </Step>

  <Step title="Click the chat icon">
    The chat icon sits in the top right corner. Clicking it opens the chat as a side panel next to your files. Drag the divider to resize it.
  </Step>

  <Step title="Prompt the agent">
    Type what you want done and send it. If you have a file open, the agent knows which one, so "summarize this file" just works.
  </Step>
</Steps>

***

## What context agents can do

<div className="qontext-cards">
  <CardGroup cols={2}>
    <Card title="Explore the repository" icon="folder-tree">
      Browse folders, read files, and search the repository by meaning or by exact text.
    </Card>

    <Card title="Create and edit" icon="pen-to-square">
      Create folders, write new files, edit existing ones, move files, and delete what's no longer needed.
    </Card>

    <Card title="Draw on synced data" icon="database">
      Search and read the raw artifacts synced from your connected sources, beyond what's already curated in the repository.
    </Card>

    <Card title="Synthesize context" icon="wand-magic-sparkles">
      Combine repository files and synced data into new context: replicate a document, aggregate scattered notes, or distill many sources into one file.
    </Card>
  </CardGroup>
</div>

Some prompts to start with:

* "Summarize everything our synced sources say about the ACME account into one file under /Customers."
* "Create a folder structure for our sales playbooks and move the existing playbook files into it."
* "Compare this file against the latest synced data and update anything that's outdated."

***

## How edits land in the repository

Context agents never bypass the repository's governance. Every edit is recorded as a change:

* It appears in the **Changes** tab with a change summary and file diffs.
* It creates a new [file version](/features/version-history), so you can always look back or restore.
* On files with **Require review** turned on, the change is held for [review](/features/change-reviews) and only applied once someone accepts it. All other changes are applied automatically.

***

## Permissions

A context agent acts with your permissions and never beyond them. It can only read files you can read and edit files you can edit; anything outside your access is invisible to it. Access to synced data is checked the same way.

<Info>
  **Permission-aware by design.**

  Two members can send the same prompt and get different results, because each agent only sees what its user is allowed to see. See [Access management](/features/access-management).
</Info>

***

## Working in a conversation

Responses stream in live, and the agent shows its work: each action it takes, like reading, searching, or writing files, appears in the transcript and can be expanded to see the details. Files the agent mentions are linked, so you can jump straight to them in the repository.

You stay in control while it runs:

* **Stop** an in-flight response at any time; the partial answer is kept.
* **Start a new chat** with the plus button in the panel header.
* **Resume past conversations** from the chat dropdown; conversations are saved per member.