dsh-vps-manager
Use your VPS inside DSH as seamlessly as over SSH: besides a curated set of common commands, you can also have the model in your conversation operate the VPS.
English | 中文
Manage your VPS from DeepSeek Harness (DSH). Check server status with commands that skip the model and cost no tokens, tell the AI what to do and let it work on the server, and handle common installs and maintenance with recipes.
Everything goes over SSH key login and shares one execution mechanism:
- Operations that change things run as remote tasks and keep going on the server if the connection drops
- Changes on the same machine run one at a time, never in parallel
- Config files are backed up before they are edited
- Before firewall or SSH changes, an automatic restore is set up on the server first
- Every run is written to an audit log
Contents
- Three ways to use it
- Requirements
- Install
- Adding a machine
- Choosing a machine in a conversation
- Terminal in the conversation
- Commands
- Talking to the AI
- Recipes
- Settings page
- What it does not protect against
- Where data lives
- Development
Three ways to use it
| Look things up | Talk to the AI | Install and maintain with recipes | |
|---|---|---|---|
| How | Commands like /vps-sysinfo |
"Install nginx on this machine and reverse-proxy a.com to port 3000" | /vps-install <recipe id> shows the plan, /vps-yes runs it |
| Uses the model | No | Yes | No |
| Cost | Zero tokens | Normal token use | Zero tokens |
| Scope | Fixed read-only queries | Anything SSH can do, with confirmations by risk level | 29 recipes that are safe to run again |
Requirements
- DeepSeek Harness 0.1.5-rc.2 (tested on DSH Desktop 2.0.10)
- macOS or Linux. The plugin relies on OpenSSH connection sharing, which the OpenSSH bundled with Windows does not support
- Node.js >= 22.13
- SSH key login only. Password login is not supported, and neither is
sudothat asks for a password - Operations that change the system need the remote user to be
rootor to have passwordlesssudo; read-only queries do not
Install
Plugin market: search for dsh-vps-manager in DSH's plugin market (dshmarket) and click install. Works in both DSH Desktop and dsh web.
Command line: in DSH Desktop, click the DSH icon in the menu bar (the system tray on Windows) → "Open DSH Terminal"; with command-line DSH (dsh web), use an ordinary terminal. Run:
dsh plugin add dsh-vps-manager
For the latest code on GitHub (possibly newer than the release on npm), use github:AIcivilization/dsh-vps-manager as the package name instead.
Restart DSH afterwards. Plugins are loaded when DSH starts.
To uninstall, open DSH Settings → VPS Manager, scroll to the bottom and click "Uninstall…" (see Settings page). You can also run dsh plugin remove dsh-vps-manager in a terminal and restart DSH; that removes only the plugin and keeps the machine list, keys, SSH configuration and audit log.
Adding a machine
Open DSH Settings → VPS Manager → "+ Add machine". The wizard has four steps:
- Details: address, port, user, alias, note, group
- Key: generates a dedicated key
~/.ssh/dsh_vps_ed25519and leaves your existing keys alone - Place the public key on the server, in any of three ways. The plugin never touches your password:
- Copy the public key and paste it into your provider's "SSH keys" page
- Copy a one-line command and run it on a server you can already log in to
- Open a terminal with
ssh-copy-idalready filled in, and type the password yourself
- Save and test the connection: the connection settings go into
~/.ssh/config.d/dsh-vps.conf, oneIncludeline is added at the very top of~/.ssh/config(backed up first to~/.ssh/config.dsh-bak), and the machine gets a health check
Machines already defined in ~/.ssh/config can be taken over with "Import from ~/.ssh/config". The plugin never rewrites what you wrote in ~/.ssh/config.
Choosing a machine in a conversation
The conversation header has a VPS switch: the word VPS, then the terminal button >_, then one small square per machine. Each square shows its number (1 for the first machine, 2 for the second, and so on), matching the numbers on the settings page.
A square's colour is the real connection state, not just whether it is selected:
- Grey: this conversation has not selected that machine
- Yellow (pulsing): selected, connecting
- Green: selected, and just checked to be reachable
- Red: selected, but unreachable. The reason is shown below the input box (for example "the SSH configuration has no entry for this machine") with a retry button; hovering the square shows it too
A live check runs when you turn the switch on, open the conversation, or come back to the DSH window (tens of milliseconds over the shared connection). Every command, AI tool call and terminal connection also records whether it reached the server, and the square follows
Click a square: this conversation enters VPS mode. From then on
/vps-commands act on that machine, and the plugin tells the model which machine this conversation operates on, what system it runs (package manager, init system, privilege), which program holds ports 80/443, which services and containers are running, plus a few rules it must follow (never guess names, handle systemd services only through systemctl, never turn a failed lookup into a restart or reinstall). The model writes commands for that system, and knows what you mean by "this machine" or "the server"In VPS mode the model cannot use local bash: a call is refused with a pointer to the VPS tools, so the model does not investigate server problems on your own computer. Reading files, searching and similar tools are unaffected
Click it again: VPS mode ends, the square turns grey, and local bash is available again. With nothing bound there is no default machine at all: commands do not run, and the AI must name the machine it operates on
Each conversation can be bound to one machine at a time. A binding only applies to its own conversation: one window can operate the server in VPS mode while another keeps working on local code, without affecting each other
The same from the keyboard:
/vps-use <alias>binds,/vps-use offunbindsA machine that has never had a health check gets one in the background when the switch is turned on, so the model receives its system details
Nothing is shown below the input box, except a single line in three cases: a task is running on the machine, the machine is unreachable, or its disk is at least 85% full.
Terminal in the conversation
The >_ button right after VPS in the conversation header is the terminal. Once a machine is bound, click it and a real terminal opens below the input box, working just like an SSH session (with only one machine you can click it before binding, and it binds that machine for you):
- Menu scripts,
top,htop,vim,docker exec -it,mysqland other programs that need keystrokes all work, as do Ctrl+C, arrow keys and CJK text - Three round buttons at the top right of the terminal:
- Red ×: end this terminal; the shell on the server and anything running in it end too
- Yellow −: minimize to a bar below the input box ("terminal running in the background · open N minutes"); the terminal keeps running, and clicking the bar brings it back
- Green: maximize; the terminal fills the conversation area and the input box moves to the top. Click again (or double-click the title bar) to restore
- The
>_button in the header is still a switch: not open → open; open → minimize; minimized → restore. While minimized it shows a small green dot - Minimizing, switching to another conversation and coming back keeps the terminal and everything on it
- Reloading the page or losing the network: the server keeps the terminal for a while (10 minutes by default, adjustable in settings). Come back within that time and it reconnects automatically, replaying the output you missed; only after that does it end
- At normal size, drag the bottom-right corner to change the height; full-screen programs redraw at the new size, and the height is remembered
- Turning the VPS switch off or moving to another machine ends that conversation's terminal
- What you type and see here does not go to the AI; run a command with
/vps-shwhen you want the AI to see its output - Opening and ending are recorded in the audit log (keystrokes are not)
Settings (DSH Settings → VPS Manager → 终端 / Terminal): colour scheme (follow system / dark / light, where follow system matches DSH's appearance), font size (11–20), how long to keep the terminal after a disconnect (5 minutes / 10 minutes / 30 minutes / 1 hour), and whether other devices may open it.
It works in both DSH Desktop and dsh web: the connection follows the page address (an https page automatically uses an encrypted connection). Every connection must pass three checks: DSH's own sign-in check, an origin that is the DSH page itself, and the plugin token embedded in that page. On top of that, by default the terminal only opens on the computer running DSH: when you reach dsh web through a LAN address or a reverse proxy, first tick "允许从其他设备打开 VPS 终端" (allow opening the VPS terminal from other devices) under DSH Settings → VPS Manager → Terminal. The terminal is full control of the server, so only turn this on for access paths you trust.
No native module has to be compiled on your machine: the pseudo-terminal on the server is requested with ssh -tt, and window-size changes are applied over a second SSH connection. The terminal display is xterm.js (MIT licensed, bundled with the plugin and loaded the first time you open a terminal).
Commands
Commands skip the model and cost no tokens. DSH folds a command's result down to one line, so the first line of every result is the conclusion.
Two rules:
- Commands act on the machine bound to the current conversation. If nothing is bound, they do not run.
- Commands listed without arguments below must be sent as the bare command name. DSH only passes the text after a command to the plugin when the command declares arguments. Add text after a command that takes none, and the whole line is sent to the model as an ordinary message. Anything that needs confirmation shows a plan first, and you confirm by sending
/vps-yeson its own.
Commands you type are shared with the AI: DSH itself never passes command results to the model. The plugin keeps your last few /vps-sh commands and their output (tokens, passwords and private keys masked first) and attaches them the next time you talk to the AI, so you can investigate yourself and then just ask "why did that fail?". Prefix --private to keep a command out of it.
Entry points
| Command | What it does |
|---|---|
/vps-help |
Every command and how to use it |
/vps-yes |
Confirms the plan you just saw: a reboot, an install, or a dangerous command that was held. Only for the current conversation, valid for 5 minutes, runs once |
Looking things up
| Command | What it does |
|---|---|
/vps-sysinfo |
OS, CPU, memory, disk, load, public IP |
/vps-disk |
Mounts, inodes, largest directories (the scan has a time limit) |
/vps-ports |
Listening ports and the processes behind them |
/vps-services |
Running and failed services |
/vps-net |
Network interfaces, cumulative traffic, connection count |
/vps-docker |
Containers and disk usage |
/vps-ping |
Hostname, OS, load, uptime |
/vps-logs <service> |
The last 100 log lines of a service |
/vps-q <recipe id> |
Runs any query recipe, e.g. ip-info, top-procs, cron-list, cert-expiry, firewall-status, updates, login-history |
/vps-sh <command> |
Runs a command on the machine and shows the output in the conversation: the directory you cd into is remembered within the conversation; commands that keep refreshing or page (top, tail -f, journalctl -f, less, watch) are turned into one-shot output, and things that cannot work here (vim, interactive shells) point you to the terminal; commands judged dangerous are held until you send /vps-yes. Prefix --bg to run in the background, --private to keep this output away from the AI |
Machines
| Command | What it does |
|---|---|
/vps-list |
Registered machines and their status; ★ marks the one bound to the current conversation |
/vps-use <alias> |
Binds the current conversation to a machine; /vps-use off removes the binding |
/vps-probe |
Runs the health check again: OS, init system, package manager, privilege, CPU, memory, disk |
/vps-reboot |
Checks before rebooting: why a reboot is needed (kernel or libc updates waiting), whether now is a safe time (it refuses while the package manager is installing or a plugin task is running), which containers will stop, and whether they will start again by themselves. /vps-yes reboots, waits for the machine to come back, and reports the kernel change, containers and failed services. If the machine is still unreachable after 5 minutes, it tells you to check your provider's console |
Installs and tasks
| Command | What it does |
|---|---|
/vps-recipes |
The recipe list, grouped into installs, configuration and queries |
/vps-install <recipe id> [key=value …] |
Shows the plan: detection result, parameters and their defaults, steps, and the script itself. /vps-yes runs it |
/vps-tasks |
Remote tasks on the machine |
/vps-task <task id> [--stop] |
A task's state and log; add --stop to terminate it |
Troubleshooting
| Command | What it does |
|---|---|
/vps-doctor |
Plugin version, the machine bound to the current conversation, a connectivity test, the most recent runs |
Talking to the AI
The plugin registers 5 tools for the model, plus a set of operating rules (the vps-operator skill, which the model reads when it needs to):
| Tool | What it does |
|---|---|
vps_hosts |
Lists registered machines |
vps_exec |
Runs a script on a machine |
vps_write_file |
Writes a remote file: backs it up first and restores it if validation fails |
vps_task |
Remote tasks: list, status, log, terminate |
vps_recipe |
Recipes: list, show, run, and save what was just done as a recipe |
Confirmation by risk level
Before any script runs, the plugin decides its risk level. The level the model declares can only raise the result, never lower it.
| Level | Examples |
|---|---|
| Read-only | df -h, systemctl status, docker ps, journalctl |
| Change | apt-get install, sed -i, systemctl restart, writing files |
| Dangerous | rm -rf, mkfs, firewall changes (ufw, iptables), passwd, reboot, killing processes (kill, pkill, killall), curl … | sh |
Whether a confirmation dialog appears depends on the machine's confirmation level (machine > group > global):
| Level | Read-only | Change | Dangerous |
|---|---|---|---|
| Careful (default) | Automatic | Ask | Ask |
| Relaxed | Automatic | Automatic | Ask |
| Fully automatic | Automatic | Automatic | Automatic |
| No approval UI available | Automatic | Refused | Refused |
Three safeguards
- Files are backed up before edits.
vps_write_filecopies the original to~/.cache/dsh-vps/backups/on the server, writes atomically, runs the validation command you give it (such asnginx -t), and restores the original if validation fails - Connectivity safety net. Before firewall, SSH or network changes, a timed restore is set up on the server (120 seconds by default). After the change, a brand-new connection is used to test access. If it cannot connect, the restore runs when the time is up, so you are not locked out
- Long operations survive disconnects. Operations that change things run as remote tasks that continue on the server if the connection drops, and you can check on them again later. Cancelling only stops waiting and does not kill the task; terminating a task is always an explicit action
Recipes
29 built-in recipes, each safe to run again:
- Installs (7): Docker, Nginx, fail2ban, common CLI tools, Portainer, Uptime Kuma, Nginx Proxy Manager
- Configuration (6): system update, system cleanup, swap, timezone, BBR congestion control, automatic security updates
- Queries (16): system info, disk, ports, services, network, containers, service logs, reachability and load, health check, public IP, top processes, scheduled tasks, certificate expiry, firewall status, available updates, login history
Install and configuration recipes have four parts:
- detect: checks whether it is already installed or already in the target state
- plan: the steps, written for people
- run: does the work
- verify: proves it actually works, e.g.
docker inforather thandocker --version
A recipe that is already in its target state skips the install and goes straight to verification.
Pass parameters as key=value. The plan lists every parameter with its default:
/vps-install setup-swap size_mb=4096
/vps-yes
Adding your own
- Ask the AI to save one. After it has done something reusable for you, say "save this as a recipe". It turns the specific values into parameters and adds
detectandverify; the plugin checks that no password or key slipped in, then saves it as$DSH_HOME/vps-manager/recipes/my-<id>.yml - Write a YAML file yourself and put it in that directory
Your own recipes get a my- id prefix and cannot replace built-in ones. To change one, save it again under the same id; to delete one, delete its file.
Your own recipes are treated as untrusted: their risk level is the stricter of what the recipe declares and what the plugin's static analysis finds, and the first run after a recipe is added or changed asks for confirmation at least at the "change" level.
Settings page
DSH Settings → VPS Manager, all point-and-click, never through the model:
Machine list: number (matching the squares in the conversation header), address, privilege, group, note; one-click connectivity test
Add machine, and Import from ~/.ssh/config
Per-machine settings: alias, address, port, user, jump host, public key placement, host fingerprint, confirmation level, removal
Basics: fill in the state you want (timezone, swap size, BBR, automatic security updates, fail2ban, common CLI tools). On save, only the items that differ from the current state are run, each as a remote task with progress shown
Global settings: default confirmation level, safety-net duration, and whether changes may be made from the settings page when DSH's web server is open to the local network
Terminal: colour scheme (follow system / dark / light), font size, how long to keep the terminal after a disconnect, and whether the terminal may be opened from other devices (off by default)
Data location
Uninstall: tick what to do, confirm, and see the result of each step. On DSH Desktop the plugin can be removed directly, followed by a one-click DSH restart; elsewhere you get the terminal command to run
- Remove the plugin itself (ticked by default)
- Remove the SSH connection settings (ticked by default): the
Includeline the plugin added at the top of~/.ssh/configis removed andconfig.d/dsh-vps.confis renamed as a backup, so both can be restored - Clean up the plugin directory
~/.cache/dsh-vpson the servers (machines with a running task are skipped) - Revoke the plugin key's login access on the servers (its line is removed from
authorized_keys, after a backup). If that key is your only way into a server, you will be locked out - Delete the plugin's dedicated key, and delete the plugin data (machine list, audit log, your own recipes); these cannot be restored
Server-side items run first, because once the local configuration and key are gone the servers can no longer be reached. Only the first two items are ticked by default
The settings page's backend only accepts same-origin JSON requests carrying a token generated at random each time DSH starts. It offers no arbitrary command execution and cannot write arbitrary files. When DSH's web server is bound to 0.0.0.0, operations that change a server are turned off by default.
What it does not protect against
- Risk-level confirmation guards against AI mistakes, not against an AI set on getting around it. Static analysis cannot recognise every disguised form, and the model can also use DSH's own bash tool to ssh into the server directly
- DSH's sandbox does not cover the ssh processes this plugin starts itself
- Nothing typed in the terminal goes through risk-level confirmation. You are typing the commands yourself, exactly as in SSH, and the plugin does not intercept them
- There is no general rollback. What you have is file backups, the connectivity safety net and recipes that can be run again. Take a snapshot in your provider's console before reinstalling the OS, upgrading to a new major release, or repartitioning
Where data lives
| Location | Contents |
|---|---|
$DSH_HOME/vps-manager/hosts.yml |
Machines, groups, confirmation levels (safe to edit by hand) |
$DSH_HOME/vps-manager/state.json |
Health check results and which machine each conversation is bound to |
$DSH_HOME/vps-manager/recipes/ |
Your own recipes |
$DSH_HOME/vps-manager/audit/ |
Audit log: one JSON line per run (source, machine, action, level, result), one file per month, kept for 6 months |
~/.ssh/config.d/dsh-vps.conf |
SSH connection settings maintained by the plugin |
~/.ssh/dsh_vps_ed25519 |
The dedicated key the plugin generates |
~/.cache/dsh-vps/ on the server |
Remote task directories, logs, file backups |
The plugin never reads private key contents, and passwords never pass through it.
Development
npm install
npm test
194 tests, no real server needed: a local sh -s stands in for the remote sshd, covering the payload protocol, remote tasks, locking, backup and restore, risk classification, VPS mode, uninstall, commands, the settings-page backend, the terminal connection (authentication, local-only access, input and output, window size, keeping and resuming after a disconnect, cleanup) and UI rendering. On a machine with DSH Desktop installed, tool definitions, return values, skill fields and approval outcomes are also checked against DSH's own dsh-tools, dsh-skill and dsh-user-approval.
License
Bundled third-party code: lib/vendor/xterm/ contains xterm.js 6.0.0 and addon-fit 0.11.0, MIT licensed, copyright the xterm.js authors; see lib/vendor/xterm/LICENSE.
No comments yet. Be the first to write one.