A gated AI diagnostic console for Linux. You describe a problem; the AI proposes commands; you decide what runs and what output the AI gets to see.
Warning
Beta. DA Toolkit is beta software and is not intended for production use. Use it at your own risk. It helps you run commands on real systems, so read every command before you run it: the risk labels and reviews are aids, not guarantees.
A training scenario: the scripted training model and a mock host, no real systems.
- The AI cannot execute anything. Its proposals land in a queue with a risk badge (read only, modifying, disruptive) and a flag for commands that may expose secrets. Only your click types a command into a terminal.
- You choose what goes back. Output is captured per command, secrets are redacted, and you review and edit the exact text before the AI sees it.
- Real terminals. Local, SSH and WinRM sessions are real PTYs (
sudo, Ctrl-C, pagers and device CLIs behave normally), and RDP opens in a tab through Apache Guacamole's guacd. - Works through ScreenConnect. A window session follows a ScreenConnect (or TeamViewer, AnyDesk, VM console) window on your screen (X11 or XWayland): the AI sees it through screenshots you send, and you paste the commands it proposes. No setup or API needed.
- Any OpenAI-compatible model: NanoGPT, Ollama, LM Studio, vLLM and others. See which models work well for how nine popular models did.
- Privacy tiers per case. Open, Confidential and Sovereign cases limit which models may see the data: anything, end-to-end encrypted or local, or local only. End-to-end encrypted and TEE models are attested before anything is sent.
- An explicit investigation: a hypothesis board the AI keeps up to date, recipes for common checks, baselines and diffs, dry runs, rollbacks and second opinions from a reviewer model before risky changes.
- A record of the case: an audit log, terminal transcripts, a Markdown export and AI-written ticket summaries, client updates and runbooks.
- Secrets live in the OS keyring (GNOME Keyring / KWallet), never in config files.
MSPs and sysadmins who work from Linux. DA Toolkit is developed and tested on Debian/Ubuntu/Mint. It manages Windows machines fine (WinRM and RDP), but it runs on Linux.
Anyone is welcome to port it to Windows or macOS, but I won't be making any effort to do so or to maintain a port.
Releases come as a single AppImage for x86-64 Linux: download it, make it executable
(chmod +x DAToolkit-*.AppImage) and run it. It needs Ubuntu 24.04 / Mint 22, Debian 13, or a
current Fedora, Arch or CachyOS (X11 or Wayland). The app window uses the system's WebKitGTK,
so that it gets your distribution's security updates:
| Distribution | For the app window | For RDP sessions (optional) |
|---|---|---|
| Ubuntu, Mint | sudo apt install gir1.2-webkit2-4.1 |
sudo apt install guacd |
| Debian | sudo apt install gir1.2-webkit2-4.1 |
not packaged: guacd container (below) |
| Fedora | sudo dnf install webkit2gtk4.1 |
sudo dnf install guacd libguac-client-rdp, then sudo systemctl enable --now guacd |
| Arch, CachyOS | sudo pacman -S webkit2gtk-4.1 |
not in the repositories: guacd container (below) |
Where guacd isn't packaged, DA Toolkit runs Apache Guacamole's own image for you: it starts the container when you open an RDP host and stops it when the last RDP session closes or the app quits, published on this machine only (guacd's link is unencrypted). Download the image once:
podman pull docker.io/guacamole/guacd:1.6.0@sha256:8974eaa9ba32f713daf311e7cc8cd7e4cdfba1edea39eed75524e78ef4b08f4f(docker pull ... works the same.) Arch's AUR has guacamole-server,
but it doesn't build as is: current glibc trips its -Werror, and its RDP plugin doesn't
compile against Arch's FreeRDP 3, so it also needs freerdp2 from the AUR. If guacd isn't running when you open an RDP host, DA Toolkit
shows the commands for your distribution in a dialog you can copy them from.
Without WebKitGTK, DA Toolkit opens in your default web browser instead and says what to install.
You can also choose the browser yourself (Settings → General, or --browser); the browser
version has a Quit button, because closing the tab doesn't end the app. Every release can
be rebuilt from its commit to the same bytes: see packaging/README.md.
To run from source instead (Python 3.11 or newer):
sudo apt install python3-venv python3-gi gir1.2-webkit2-4.1 # Debian/Ubuntu/Mint
git clone https://github.com/DigitalArchon/DAToolkit.git && cd DAToolkit
python3 -m venv --system-site-packages .venv # system-site-packages gives access to GTK/WebKit
.venv/bin/pip install -e .
.venv/bin/datoolkit--no-open only prints the local URL, for opening it yourself. Treat that URL like a password:
it grants terminal access.
- Start a case: give it a name or ticket number and pick a sensitivity. To set things up first, choose Not now: settings, providers, models and hosts don't need a case.
- Add a model under Settings → AI providers. To try DA Toolkit without an API key, add the training provider: a scripted model that walks through a canned case (the one in the screenshot) with nothing leaving the machine.
- + Session opens a local shell, a saved host (add hosts under Settings → Hosts), or a window on your screen, such as a ScreenConnect control window.
- Describe the problem in the chat, then Run, Skip and Send results as the AI works through it.
The manual covers everything else: the workflow in detail, model tiers and how attestation works, SSH/WinRM/RDP and window-session specifics, where files are stored and how untrusted output is handled.
- Connectors: ConnectWise (push ticket notes and transcripts) and Confluence (pull site notes, push runbooks)
- Newer guacd (1.5+) so guacd itself enforces the pinned RDP certificate; RDP session recording into the case timeline
- IP/hostname pseudonymisation with local reverse mapping
- Per-client notes library
See CONTRIBUTING.md. Please report vulnerabilities privately as described in SECURITY.md, not in public issues.
AGPL-3.0-or-later. See LICENSE. Vendored third-party libraries and their licences are listed in src/datoolkit/web/vendor/NOTICE.