Intelligence.
Built to outperform.

An AI agent that works in your project folders.

What it is

A desktop app for macOS, Windows and Linux. You chat with an AI agent in a window, and in the folder you choose it reads and edits files, runs commands and tells you what it did.

What it is 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 check every morning.

How it works

The app carries AgenC Core, the agent runtime. Core sends your messages to the model you pick, runs the agent's tools on your computer, and asks before it acts unless you choose a mode that lets it run. Your sessions are kept on your computer.

Your model, your key.

Nothing is locked to one vendor. Sign in with an account you already have, paste an API key, or run a model on your own machine with neither. Keys are encrypted with your system's secure storage.

• AgenC hosted models
• OpenAI
• xAI Grok
• Meta
• QwenCloud
• OpenRouter
• Ollama, on your machine

More than twenty cloud providers are listed in the app. AgenC hosted models need an AgenC account and use its credits. Each session keeps its own model and its own reasoning effort, and the model browser tells you what each one can do before you pick it: how much context it holds, whether it reasons in steps, whether it reads images. Sub-agents can run on other providers you allow. You can also talk instead of typing. The microphone in the composer transcribes on your machine with Whisper.

Talk instead of typing.

The microphone in the composer turns what you say into text. It runs on your own machine with Whisper, so nothing you say is sent anywhere to be transcribed, not to us and not to a provider.

Set up once

The first run downloads the base model, around 141 MB, so voice works from the start. A larger model is there when you want better accuracy, and you can switch between them later.

It stops when you do

The app will not sit recording in the background. Hide the window and it stops listening. Voice has its own settings, next to providers and connections.

Three ways to start a task.

Type a message and the agent works. Or pick how it should approach the job before it starts.

Goal

Implement it, prove it, have it reviewed. The change is made in a separate Git worktree, the project's own check has to pass when it has one, and an independent reviewer reads the result, within the spend and time limits you set.

Swarm

Split the work across agents. Use it when a job has parts that do not depend on each other.

Plan mode

Explore, ask, then plan before touching anything. For real work, nothing is written until you approve the plan.

You decide what it may do.

Every action the agent takes goes through a permission mode. Pick one as the default for new chats, and change it in any chat.

Ask for every tool

The agent asks before its tools run. You answer Allow once, Allow for session or Deny.

Plan without changes

It reads and plans, but does not edit.

Approve edits automatically

Edits to files in the project go through. Most commands still ask.

Bypass permissions

No questions. Core still refuses a short list of protected actions, and keeps its sandbox where the system has one.

On Windows, Core has no sandbox yet, so Bypass permissions is the only mode there.

A browser, a terminal, and answers you can use.

A browser and a terminal open next to the chat, and answers can carry charts, tables and files.

Browser and terminal

Both open next to the chat. Use them yourself, or let the agent drive them while you watch. On Windows the agent cannot drive them yet.

Run buttons

A shell command in an answer gets a Run button that runs it in that terminal, so a password or a key is typed there, never in the chat.

Charts, tables and files

Answers can carry charts, price charts, tables, equations, diagrams, images and files, and links get previews.

Git

The project's pull requests and their checks sit under the chat, and each new session can get its own worktree and branch.

Teach it the work you actually do.

The agent is general. Your work is not. Two things close that gap, and both come with a catalogue you can browse inside the app.

Plugins

Ready-made tools you install in a click. Forge checks code structure, Ledger inspects a wallet without signing, IoT Builder uploads PlatformIO projects, LLM Checker finds which local models your machine can run, Inbox sorts your mail, Quill writes in a voice you pick. You can also add your own from a local folder or a Git repository.

Skills

Instructions the agent loads when a task matches. Some come with the app, some you write, some are recommended for what you already do. They turn a repeated way of working into something the agent just knows.

Work that repeats, and work you walk away from.

Some jobs should happen without you asking, and some you want to check on from somewhere else.

Routines

Tasks on a schedule, or started by hand. It comes with a code briefing every weekday, a repository health check, a dependency review on Mondays, a release notes draft and a documentation check. Describe your own in a sentence, or, on macOS and Linux, ask the agent in a chat to set one up. Routines run while your computer is awake and AgenC is open.

Connections

Open a session from a browser on another device. Every new device needs your approval, and you choose which sessions are shared. There is an SSH tab for machines you reach that way, and a switch that stops the computer sleeping while connections are on.

Telegram agents

Give an agent its own role, its own folder and its own Telegram bot, linked to your Telegram account, then message it like a person. Each one runs on your computer and answers from there.