DocsAgenC Desktop

Desktop overview

What AgenC Desktop is, what it is for, how it works with AgenC Core, and where each feature is documented.

AgenC Desktop is an app for macOS, Windows and Linux where an AI agent works on your computer. You chat with it in a window. In the folder you choose, it reads and edits files, runs commands, uses a browser and a terminal, and tells you what it did.

Use it for work you would hand to a capable assistant: build a feature, fix a bug, review a repository, research a question, or run the same job every morning. You stay in charge of what the agent may do without asking.

Inside, the app carries AgenC Core, the agent runtime of AgenC. Core sends your messages to the model you pick, from the providers you connect: a hosted model, a model you pay for with your own key or account, or a model on your computer. Core runs the agent's tools in your folder, asks you before it acts when your permission mode says so, and keeps every session so you can come back to it. The app is the window onto Core: the chat, the approvals, the browser and terminal, and the settings.

AgenC Desktop is in beta. Expect rough edges, and report them with Feedback in the app's account menu.

Get started

  1. Install it. One command in a terminal, on any of the three systems. See Install and update.
  2. Let the first run finish. It prepares Core and downloads the speech model for voice input. See First run and startup.
  3. Connect a model. Download a local model, sign in to AgenC for hosted models, paste an API key, or sign in with X to use Grok. See Providers and models.
  4. Open a project and say what you need. Pick a folder in the composer, type your request and press Enter. See Writing messages.

What happens when you send a message

The agent reads your request, decides what to do, and uses tools: it reads files, edits them, runs commands, searches, browses. While it works, a line under the last message says what it is doing. When a tool needs your approval, a sheet asks Allow once, Allow for session or Deny. How often it asks depends on the session's permission mode, from "Ask for every tool" to "Bypass permissions". See Permissions and approvals.

When the agent is done it answers in the chat, with text, code, tables, charts or images, and you continue the conversation. Messages you send while it works wait in a queue.

Everything the app does

Sessions and chat:

  • Sessions and projects: the sidebar, projects, pinning, search, activity, archive, and sessions that survive restarts.
  • Writing messages: models, effort and speed, attachments, the @ and / menus, queued messages and the Run button.
  • Answers in the chat: charts, tables, equations, diagrams, images, files and link previews.
  • Whisper voice input: talk instead of typing, with speech recognized on your computer.

How the agent works:

Beyond one chat:

The app itself:

Platforms

The app runs on macOS 13.5 or later, Windows 10 and 11, and 64-bit Intel or AMD Linux with glibc 2.38 or newer. Most of it works the same on all three. On Windows, where Core has no sandbox yet, the only permission mode is Bypass permissions, and the agent does not get the app's own tools: it cannot drive the browser and terminal or manage routines from a chat. Windows also needs Git for Windows. How new versions arrive differs on each system; see Install and update.

AgenC Desktop and AgenC Core

AgenC Core is also a command line tool, for a terminal, a CI job or a server. The app carries its own copy of Core and keeps its own data, so a separate Core install and the app do not share sessions, settings or keys. Core's own commands and configuration are in the AgenC Core guide.