Documentation

ProxYard is a Proxmox VE client for the desktop. It talks to the PVE API of your server or cluster with your API token or your login and your rights; nothing is installed on the server. Every action says what it does before it runs.

ProxYard is in early beta. If you find a bug or have an idea, please write to info@lixsl.net.

What it supports

Installation

Download the package for your system from the download page: a .dmg for macOS, an installer for Windows, an AppImage or .deb for Linux. The builds are not signed yet. On macOS they are not notarized: open the app with a right click, then Open. On Windows SmartScreen may warn: More info, Run anyway. To check a file, compare its SHA-256 (shasum -a 256 <file> on macOS and Linux, Get-FileHash <file> in PowerShell) with the list on the download page.

On the first start ProxYard asks for the theme and then for a server. To look around first, open the demo datacenter: a simulated cluster of three nodes, nothing is sent anywhere. After that ProxYard opens the last server by itself. The settings are kept in one file: ~/Library/Application Support/ProxYard on macOS, %LOCALAPPDATA%\ProxYard on Windows, ~/.config/proxyard on Linux.

Servers and logins

Add a server with its address (host name or IP, :port when it is not 8006). For a cluster one node is enough: ProxYard finds the others and moves to the next one when a node does not answer. Log in with an API token (user@realm!tokenid and its secret) or with user and password; the password can be remembered or asked at every start. Secrets are kept in the keychain of your computer (macOS Keychain, Windows DPAPI, the Secret Service on Linux), never in the settings file.

Switch servers with the menu in the top right, where you also add, edit and remove them. A server can be locked to read only: then every action is off for it, no start, stop, migrate, delete or config change. Reading, charts, logs and consoles to read still work. ProxYard also reads your PVE permissions and turns off what your role may not do.

The pages

The tree on the left holds the datacenter: by node (server view), by folder, by pool or by tag, with the tags in their PVE colors. Above it is the dashboard, below it the map and Proxbit. A click opens the page of a guest, node or storage; its sections are tabs at the top, and tabs that do not fit go into More.

• Dashboard. Nodes, guests, CPU, memory, storage and failed tasks of the whole datacenter, what needs attention, the busiest guests and the last tasks, on one screen.
• Datacenter. All guests and nodes as tables, storage, backup jobs and their coverage, tasks, permissions (users, groups, tokens, roles, realms, second factors), firewall, cluster and join, HA, replication, Ceph, resource mappings, SDN, options, notifications, metric servers, ACME and support.
• Node. Summary with history, shell, system (network, DNS, time, services, syslog, certificates), updates, disks (LVM, ZFS, directories), PCI devices and GPUs, tasks, firewall, Ceph, replication and subscription.
• Virtual machine. Summary with live usage and PVE's history from an hour to a year, console, hardware, options, snapshots, backups, cloud-init, firewall, tasks, replication and permissions.
• Container. The same with resources, network and DNS instead of hardware and cloud-init.
• Storage. Usage, configuration, content by kind, the biggest volumes, the guests on it, and its content: upload, download from a URL, restore and delete.
• Panel. The tasks of the datacenter (a click opens the log) and the cluster log, at the bottom. It starts as one row; click it or drag it up.

Create a VM or a container, or restore a backup, with Create in the title bar.

Consoles

Consoles open on the guest's page and connect when you open the tab. Virtual machines get the graphical console (noVNC), also for Windows guests, with keys like Ctrl Alt Del, paste and scaling. Containers, node shells and serial ports get a terminal (xterm.js). The connection goes through the app on your computer, with the same login as the rest of ProxYard.

GPUs

The GPU tab of a node (PCI and GPUs) and of the datacenter (Hardware) lists every GPU with its vendor, IOMMU group, driver and who uses it: passed through to a VM with vfio-pci or used by the host. It warns about what would stop a VM with passthrough from starting, shows the resource mappings and assigns a GPU to a VM. Raw PCI passthrough needs root@pam with a password; with an API token use a resource mapping.

Proxbit

Proxbit is the assistant. It runs on your local Claude Code (claude -p) with ProxYard as its only tool server. It reads the datacenter: guests and their config, nodes, storage, PVE's history, tasks and their logs, the syslog, backups, the firewall, users and tokens, GPUs and the network. It answers with what you need to see: KPI tiles, live cards, charts over time, tables, graphs, timelines and the city. Type / for commands like /diagnose, /capacity, /backup or /security.

It can not act. To change something, Proxbit proposes a change; you see it and accept it, or not. It never applies anything itself. One answer may call 32 tools and read 20 times at most, so a confused model can not keep the server busy.

Claude Code gets no built-in tools (no shell, no files, no web), no other tool servers, none of your Claude settings, hooks, CLAUDE.md or memory, and only the environment it needs to log in. Passwords, tokens and keys in configs, notes and logs are hidden before the model sees them.

Text on a server can try to talk to the model, for example a guest's notes or a log line that says "ignore your rules". ProxYard marks everything it reads as data, Proxbit is told that such text is never an instruction, and it only proposes changes you asked for.

The tool server listens on 127.0.0.1 only. Every claude process gets its own key, in a file only you can read, and the key ends with the process. Requests without the key, from a web page, to another host name, too big or too many are refused.

What leaves your computer: your questions and what Proxbit reads (configs, tasks, log lines, metrics, without the hidden values) go to Anthropic through Claude Code, under your Claude account and its terms. ProxYard itself sends nothing. If the data of a server must not leave your computer, do not use Proxbit on it.

Datacenter map

The map draws the datacenter as a city: guests are houses on plots, grouped by tag, pool or node, their doors open while they run, nodes are power plants by the river, and trucks carry the guests' network traffic. Drag to move, scroll to zoom and click on anything to see its details, open its page or start and stop it. The tools on the right zoom, show the whole city and switch the roofs and the time of day.

Keyboard

Security

ProxYard only does what you confirm. Actions say what they do to which guest on which node, and deleting a guest needs its VMID typed. Passwords and token secrets stay in the keychain of your computer, and a self-signed certificate is pinned by its fingerprint.

The app's own server listens on 127.0.0.1 only and takes requests only from its window. The only request to proxyard.net is the update check, once an hour, which can be turned off in the settings. Nothing about your servers is sent to us; what Proxbit sends to Anthropic is in Proxbit.

Releases

Every release is tested before it ships: the unit tests and a check of all NuGet packages for known vulnerabilities. Every night the packages of the last three releases are checked again; the security page shows the result. What changed is in the changelog.