AJEAN

English · Français

Run your own AI at home. AJEAN is a single program that turns a computer with a graphics card (or a plain CPU) into a complete AI assistant: a chat interface, memory, tools, web access, scheduled tasks, and access from your phone, with nothing leaving your machine.

AJEAN handles everything around the model. The model itself runs on llama.cpp, which AJEAN installs and keeps up to date for you.

• One file, no dependencies. No Docker, no Python, no runtime to install. Linux, Windows and macOS.
• An assistant, not just a model. It remembers, organizes work into projects, runs commands, reads the web, drives a browser, and works on its own on a schedule.
• Your data stays yours. Everything runs on your machine. Memory and conversations can be encrypted at rest, and remote access is encrypted end to end.
• Use it from anywhere. The same conversations on your computer and your phone, live.
• Plugs into your tools. An OpenAI-compatible endpoint for any third-party app.

Contents

• Install
• Using AJEAN
• What the AI can do
• Models and engine
• Access from anywhere
• Your data
• Reference
• Build from source

Install

Download the file for your system from the latest release:

Windows

Double-click ajean-windows.exe. AJEAN installs itself, adds its shortcuts, and opens in its own window, with an icon in the notification area (Open AJEAN / Quit). Running a newer file later updates the installed copy.

Then, in the interface:

1. Engine section: install llama.cpp. The pre-built version is ready in about two minutes; the compiled version is tuned for your machine but takes longer.
2. Presets section: create a preset, pick or download a model (any direct .gguf link, for example from Hugging Face).

macOS

Unzip, drag AJEAN.app into Applications, then right-click and Open the first time (the app is only ad-hoc signed). AJEAN opens in its own window and lives in the menu bar. Then install the engine and add a model as on Windows.

For command-line use, take the bare binary ajean-macos-arm / ajean-macos instead of the app.

Linux (server)

curl -L -o ajean https://github.com/nathaninline/ajean/releases/latest/download/ajean-linux
chmod +x ajean && sudo mv ajean /usr/local/bin/ajean

sudo ajean install        # services, permissions, folders
ajean llamacpp install    # builds llama.cpp for the GPU present
ajean edit                # set MODEL=/path/to/model.gguf
ajean start               # starts the model
ajean ui start            # interface at http://<host>:8090

Building llama.cpp needs git, cmake and your GPU toolkit (CUDA for NVIDIA, ROCm or the Vulkan SDK for AMD, the Vulkan SDK for Intel). AJEAN installs what it can by itself, and tells you what to do when it cannot. To force a backend: ajean llamacpp install --backend=vulkan (or cuda, hip, cpu).

On Linux, AJEAN runs as two services: ajean-engine runs the model, ajean-ui serves the interface, remote access and the OpenAI endpoint. Restarting the interface is instant and never reloads the model.

Updating

ajean update

Or the update button in the interface. Each release is verified against its published checksums before it is installed.

Using AJEAN

Four chat modes

Pick a mode from the button on the left of the input bar.

A conversation keeps the mode it started in. Tools only exist when agent mode is on (ajean agent on, or the switch in the interface); without it, every mode behaves like Base model.

Projects

A project is a workspace with its own memory and its own conversations, isolated from the others: one for your mailbox automation, one for a piece of software, one for your notes. Switch project from the bubble in the input bar.

Each project has:

• A memory: Markdown pages the AI reads and writes across conversations, with an index kept up to date automatically.
• Trackers: dated data that keeps growing (a weekly subscriber count, a weight, a revenue figure). The AI browses them level by level instead of loading everything.
• A description, given to the AI at the start of every conversation (context, constraints, tone).
• Options in its ⋯ menu: name and description, its memory, its trackers.

Jean, personal assistant (beta)

Jean is an assistant that remembers you over time, in a single conversation that never ends. Open it from the mode menu: it gets a full-screen view with its avatar.

• Its own memory: a short profile (name, family, preferences, habits), fiches (procedures, recipes, guides kept in full), and a journal where every exchange is logged and searchable. The Jean memory window lets you review, edit or delete all of it.
• Reminders and tasks: ask Jean to remind you of something, and the reminder arrives as a message from Jean, with a notification.
• A separate workspace: Jean cannot modify your projects' scripts, files or memory. It can read them, to see how something was done, and rebuild what it needs in its own space.
• A fresh start after a break: after 3 hours without a message, the model's context starts over. The thread on screen and the journal stay. Set JEAN_IDLE_HOURS to change the delay (0 = never).

Jean is a beta: not everything has been tested yet. For serious work, keep using Project or Quick mode.

History

The clock icon at the top of the side menu switches to the history. It follows what you are doing: the active project's conversations in Project mode, the quick conversations in Quick mode, the Base model conversations in Base model mode. Search covers every conversation. Conversations can be renamed, pinned as favorites, moved to another project, and exported to Markdown or JSON.

Conversations live on the server, so they are the same on every device, live: start a question on your computer, read the answer on your phone. Closing the browser never stops a reply.

In the terminal

ajean chat

A lightweight chat in the terminal, independent from the web interface, with three tools (bash, write, edit) that work in the folder where you run it. Handy for working on a project directory.

What the AI can do

These tools are given to the model only when agent mode is on.

Terminal and files. The AI runs commands (bash, or cmd.exe on Windows) and writes or edits files. It works in a disposable workspace; scripts worth keeping go into a separate scripts folder that a cleanup never touches.

Memory. Per project, four settings: injected (the index is loaded up front), search (the AI looks things up when needed, lighter on context), on demand (only when you ask), or off.

Web. web_search, web_open, web_read and web_grep. The built-in engine works out of the box. For pages that need JavaScript, connect a Crawl4AI server you host:

ajean internet on                          # built-in engine
ajean internet engine crawl4ai             # or a Crawl4AI server
ajean internet url http://localhost:11235

Browser control. With ajean computer on, the AI drives a real Chrome, Chromium or Edge on the machine: open a page, click, type, scroll. It targets elements through the accessibility tree, so it works even with small text-only models. You see a live preview while it browses.

Images. Attach an image to a message, and a multimodal model sees it and can use it. Images are resized before being sent.

Files. Send files to the AI from the chat (up to 1 GB each), and download the files it creates with a click.

MCP servers. Plug in any Model Context Protocol server (files, databases, mail, APIs) from the interface, over stdio or HTTP. The format is the one used by Claude Desktop, so an existing configuration copies over as is. Servers and individual tools can be switched on and off.

Scheduled tasks. The AI works on its own, at the frequency you choose (every N minutes, hours or days, or a cron expression): check a mailbox, summarize news, follow a figure. A task can also be a script only, run without loading the model. Each task chooses its project, its preset, and whether it may use memory and the web. A master switch pauses everything.

Notifications. Get notified when a reply is ready, even with the app closed or the phone locked. On iPhone, add AJEAN to the home screen first.

Long conversations. When the context fills up, older turns are summarized automatically. Long tool results are archived and can be recalled by the AI when needed, so nothing is truly lost.

Models and engine

Presets

A preset is a complete model configuration: model file, context size, GPU layers, KV cache, sampling, reasoning, speculative decoding, vision. Switching preset reloads the model, from the interface or with ajean switch. Presets are edited in the interface; `.gg