AJEAN

English · Français

Your AI models run at home, in a single binary: chat, persistent memory, web access, tools, encryption at rest, and remote access encrypted end to end.

AJEAN provides everything around the model: the chat interface, the assistant's tools, service management, and hardware management. The inference engine is llama.cpp, which AJEAN compiles itself for the machine it runs on.

download the binary  →  ajean llamacpp install  →  ajean edit  →  ajean start  →  you're live

No runtime dependency, no CMake flag to remember, no container. You get a full chat interface and an OpenAI-compatible endpoint for your third-party tools.

What AJEAN does

An assistant, not just a model. The web interface offers chat with visible reasoning, persistent memory, automatic context compaction as the conversation grows, multiple saved sessions, an editable system prompt, and appearance settings synced across devices.

Real tools. ajean agent on turns on, in one move, all of the model's capabilities on the machine:

It sees and it drives. The AI can receive images in chat (attachments, screenshots) and analyze them when the model is multimodal. With ajean computer on, it drives a real browser on the machine (open a page, click, type, scroll) to carry out tasks on the web.

Hardware and engine, handled for you. ajean llamacpp install clones and compiles llama.cpp with the right flags for this machine: CUDA (compute capability detected per GPU, so multi-GPU works), ROCm, Metal, Vulkan, or a CPU fallback. ajean llamacpp update fetches the latest commit, stops the service while it recompiles, then restarts it.

Services, not scripts. ajean install writes the two systemd units, the sudoers rules, and the folders. Then start, stop, status, logs. Windows and macOS have their native equivalents (see below).

Several models, one click. Presets each keep their full configuration: switching from one to another reloads the model without touching a file. The preset editor also covers sampling, the KV cache, flash attention, speculative decoding, and reasoning mode. .gguf files can live on any disk.

Your data stays yours. Optional encryption at rest protects memory and conversations with a key that only lives on your devices. Notifications tell you when a reply is ready, even with the app closed. The scheduler runs recurring tasks on its own.

Reachable from anywhere. ajean link opens an outbound connection to ajean.link, so no port to open, and it works even behind CGNAT. Chat is encrypted end to end: the relay never sees the conversations.

Getting started

1. The binary

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

Published binaries: ajean-linux, ajean-linux-arm, ajean-macos, ajean-macos-arm, ajean-windows.exe, ajean-windows-arm.exe. The -arm suffix means arm64, no suffix means x86-64.

2. Install and compile the engine

sudo ajean install        # two systemd units, sudoers, folders
ajean llamacpp install    # compiles llama.cpp for the GPU present

Requires git and cmake, plus the accelerator toolkit (CUDA, ROCm...) for GPU acceleration.

3. Start

ajean edit      # set MODEL=/path/to/the-model.gguf
ajean start     # starts the engine (ajean-engine)
ajean test      # check the model responds
ajean ui start  # starts the interface (ajean-ui) at http://<host>:8090

AJEAN runs as two services: ajean-engine, which runs the model, and ajean-ui, which serves the web interface, the remote-access tunnel, and the OpenAI endpoint. Splitting them lets you restart the interface, which is instant, without reloading tens of gigabytes of model.

Commands

Engine (ajean-engine):
  start | stop | restart        manage the service
  status | logs                 state / live logs
  enable | disable              start on boot
  edit                          edit the configuration in $EDITOR
  switch [N]                    change preset (presets/)
  test | bench [N]              check the model responds / measure tok/s
  vram | gpu [index...]         VRAM / GPU selection (gpu all = all)
  set-api-key [key]             protect the inference engine (Bearer)
  network [on|off|status]       make the OpenAI endpoint reachable on the LAN
  llamacpp install|update|status

Interface (ajean-ui):
  ui [start|stop|restart|status]  drive the interface service
  web [PORT]                    serve the interface in the foreground (default :8090)
  set-web-key [key]             protect the control API

Interaction:
  chat [system-prompt]          chat in the terminal
  export [options] [file]       export the conversation (Markdown, --json, --last N...)
  agent [on|off|status]         turn on ALL tools (terminal, files, memory)
  computer [on|off|status]      browser control (the AI drives a local Chrome)
  memory [off|ondemand|always]  memory mode
  internet [on|off|engine <go|crawl4ai>|url <url>|key <key>]   web access

Remote access (ajean.link):
  link <token>                  registers the token and opens the tunnel
  link code                     pairing code (10 min, single use)
  link status | logout

Installation:
  install | uninstall
  update [--check]              update from GitHub releases
  where | version

Configuration

Everything lives under $AJEAN_HOME (/etc/ajean on Linux/macOS, %ProgramData%\ajean on Windows):

Added to these at the root are the few files that cannot go elsewhere: .e2e_key (private key for end-to-end encryption), certs/ (TLS certificates managed by certmagic), and the services' logs and PID files.

The database, a single bbolt file, replaces the dozen state files of old. What you read and edit by hand stays as files: the presets, the memory pages, and, of course, the models.

The engine configuration is edited with ajean edit, which lays it out as key=value in $EDITOR:

The API key (ajean set-api-key) is stored outside the configuration, so it survives preset switches.

Models on another disk. .gguf files do not have to live in $AJEAN_HOME/models: in the interface's preset editor, under Model, Model folders, add the folder you want. Its models appear in the list, grouped by folder. The list is saved in the database, so it is kept across presets.

Environment variables

The AI's capabilities

Sessions and memory

Each conversation is a persistent session with a stable identifier. The Sessions button lists all kept conversations, you reopen one with a click (the current one is first saved into its own), and new session starts a blank thread. Sessions can be favorited and are kept in the database, so they are shared across every device connected to the same server.

Beyond sessions, the AI keeps Markdown pages under $AJEAN_HOME/memory/, re-read and updated across conversations. Three modes, independent of agent mode:

ajean memory always     # (default) it searches before answering and saves on its own
ajean memory ondemand   # tools available, but used only on request
ajean memory off        # memory off

Encryption at rest

Optional, off by default. A single switch in the settings (*enable e