How to Keep DeepSeek Harness Private on VPS
Install DeepSeek Harness on Linux VPS, pin the npm version, see wetin plugins fit do, and open port 3080 web UI safely through SSH tunnel.
Wetin be DeepSeek Harness
DeepSeek Harness (dsh) na Node.js agent runtime wey you fit run for VPS (virtual private server). Di correct way to run am na make you bind am to 127.0.0.1, den use SSH (secure shell) tunnel take access am from your browser. E dey serve web UI (user interface) for port 3080 instead of make e just stay inside terminal. For di 0.1.0-rc.6 release wey dis guide dey use, dat web server no dey ask for password. So, if you publish port 3080, anybody wey find am go get access to agent wey fit read your files and run commands as your Linux user. Newer releases, from 0.2.0-rc.2 go reach 0.2.2-rc.2 of 11 November 2026, wey be di newest release wey GitHub list for 12 November 2026, dey print startup URL wey get fresh token for dat process. Dem dey swap dat token for session cookie, but dem still dey listen on loopback by default and dem still dey refuse to bind every interface.
DeepSeek publish am under MIT licence, as npm package @deepseek-ai/dsh. Di project describe imsef as developer preview and dem talk say make you expect changes wey fit break compatibility. Di version wey we pin for down here na di August 2026 release. For 12 November 2026, npm latest tag dey point to 0.2.2-rc.1. GitHub list di newer prerelease 0.2.2-rc.2, wey dem tag as dsh-v0.2.2-rc.2 and release for 11 November 2026. Check di repository before you copy any of am enter server wey get important data.
One idea dey run through di whole design: everything na plugin. Di model adapter, tool registry, session log, sandbox, scheduler, and di agent loop itself, all of dem na plugins wey dem load into one shared context, and you fit replace any of dem. No privileged core dey wey plugins just dey decorate. Na dis one make di harness worth to try, and na also here di real risk dey. Whether dat trade-off make sense depend on wetin you dey use measure am, and how e take stand against Claude Code and Omnigent compare di plugin-everything design wit two other answers on model coupling, licence, and how much VPS resources each one dey ask for.
Harness no be model
Na di harness dey run di agent loop. Di reasoning dey happen inside model wey dey somewhere else, so nothing go work until you give am either API (application programming interface) key or di address of one model endpoint wey you host by yourself. Everything wey dey inside this post na harness configuration, e no be model behaviour, and di line wey divide di two dey important make you understand am before you spend afternoon dey find whether na which side di problem come from.
You go configure dat one for UI under Settings come select Models. DeepSeek get im own card, and for current releases, di "Add model provider" button wey dey under am get two modes. "Third-party model provider" cover di providers wey di harness don already get catalog for, like OpenAI and Anthropic, where you go paste key and di catalog go provide di endpoint and model list. "Custom model API" na di interesting option: e dey take lowercase provider ID, display name, base URL, API protocol, credential, and at least one model. Di protocol picker get OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Gemini API, so you fit configure gateway or local server wey fit talk one of those; endpoints wey get difference for request-format need compatibility settings. Custom providers fit also query di OpenAI-compatible GET /models endpoint to fill di model list for you.
Na so you go point di harness go model wey dey di same VPS. Ollama dey expose OpenAI-compatible API for http://127.0.0.1:11434/v1/, and di API key field dey take any string, ollama as standard, because di OpenAI-style client need value and Ollama no dey care about am. Ollama own docs sef describe ollama launch dsh, wey dey install @deepseek-ai/dsh if e no dey and e dey start am against Ollama model, so check which version e dey install before you rely on di pin wey dey below. Whether model wey small reach to fit your VPS good reach to drive agent na di harder question, and di difference between Ollama and vLLM as local model server go decide how much of your RAM di answer go cost.
Keys wey you type inside UI na write-only. Di harness dey store dem inside $DSH_HOME/.credentials.yaml, and di settings dey keep only credential reference, wey current releases dey write go di active profile patch file, $DSH_HOME/profiles/<profile>/cordis.patch.yml (di profile na web for dsh web). $DSH_HOME defaults to ~/.dsh. Treat di credentials file like password file, because na wetin e be: anybody wey read am fit spend your API budget. If you want edit those files directly instead of clicking through Settings, di walkthrough of dsh config files, keys and model endpoints go show you wetin each key dey do and wetin dey commot your box for each mode.
Wetin you need before you install
- one VPS wey dey run Ubuntu 24.04 or any other current Linux, wey you get SSH access to
- Node.js 24.9 or newer, because the project main branch and
0.2.2-alpha.2now declare>=24.9.0, and e get CI tests for Node 24 with compatibility check for Node 24.9 and Node 26, but no job for Node 22; tagged releases through0.2.1-rc.4still list Node 22.19 as okay, but that one no be the minimum wey the project dey build against again - one normal user account, no be
root, because the agent dey run shell commands as whoever start the process pnpmfor your PATH if you plan to install plugins, because the plugin command dey call am for shell- port 3080 make sure e close for your firewall and for your provider separate network firewall
The nodejs package wey Ubuntu 24.04 get dey older pass wetin the harness need, so for that release, install Node from NodeSource or nvm instead of using apt install nodejs. The official repository for Ubuntu 26.10 list Node 24.19.0, wey pass the Node 24.9 minimum. If the VPS dey fresh, make you harden SSH before you do anything else go take you ten minutes, because the tunnel wey you wan depend on go strong reach the SSH server wey dey behind am.
Install DeepSeek Harness for VPS, lock am to one version
node --version
npx @deepseek-ai/dsh@0.1.0-rc.6 webnpx go download the package come run the dsh binary. web na alias for --profile web, wey dey start the browser application, and the process go print the address wey e dey listen on. The default na http://127.0.0.1:3080. Those numbers na binding decision, no be just decoration, and why dsh prints a loopback address explain wetin the harness go answer and wetin e no go answer before you start find am from your laptop.
Lock the version. npx @deepseek-ai/dsh web go resolve whatever latest tag dey point to the moment you run am, except if the project wey you dey run am from don get copy installed already; if e get, npx go just use that copy quietly. Either way, no be you dey choose the version, and the project don release plenty release candidates and dem talk say breaking changes dey come. 0.1.0-rc.6 na the August 2026 version wey we lock here. For 12 November 2026, latest dey point to 0.2.2-rc.1, one release wey dey print tokenized startup URL; GitHub list the newer prerelease 0.2.2-rc.2, so to move go another version na the kind upgrade wey you suppose do on purpose after you don read the repository. Pinned version mean say the box wey you set up today go behave the same way next month, so upgrade go be decision wey you make, no be accident wey you go just discover. When pinned command still start wrong build or e refuse to install at all, the usual dsh install and version errors cover how to clear npx cache and check which npm your Node dey use.
For daily use, install am once instead of make you dey re-resolve am every time you start.
npm install -g @deepseek-ai/dsh@0.1.0-rc.6
dsh --profile web --helpThat second line dey important to run, because the launcher and the web app get different flag sets. dsh --help go show the launcher own options. dsh --profile web --help go show the flags wey the web application dey accept, na there --port, --host and the repeatable --trusted-host dey live.
Now confirm wetin e dey listen on.
ss -tlnp | grep 3080The local address column suppose read 127.0.0.1:3080. If e read 0.0.0.0:3080, the UI dey listen on every interface, so na only your firewall dey stand between am and the internet, and you suppose stop the process before you do any other thing.
Why you must never publish port 3080
The web server for 0.1.0-rc.6 no get any authentication layer. The configuration just get listen host and listen port, and na only that one dey. Access control for deployments wey no be loopback na separate trusted-host setting, and that one no be login screen.
Newer releases don add one. From 0.2.0-rc.2 go front, startup dey print URL wey get fresh token for that process; the browser wey open am go receive signed session cookie and dem go redirect am go the same address without the token. One Host/Origin fence for front of the /api routes dey accept only loopback, the listener own bind address, and the authorities wey you name with --trusted-host. That one na real credential, but the project own docs warn say the printed URL na credential wey you suppose share only with people wey you trust, and say plain HTTP dey send am without encryption. Since 0.2.1-alpha.2, the web app fit also bind one concrete local address and serve HTTPS by itself with --tls-cert and --tls-key. That one solve the plaintext wahala, but e no solve the exposure wahala. E just make the exposed port harder to abuse. E no mean say you suppose publish am.
Now, make we look wetin dey behind that port. The agent dey edit files for the workspace and dey run shell commands, and your provider credentials dey for disk near am. So for the pinned release, one open port 3080 na remote shell with chat interface, wey dey run as the user wey start am, with your API key attached. Nobody need exploit for that one. Dem just need the port number. For release wey get launch token, dem still need the token, wey anybody wey fit read your terminal, your service logs, or your unencrypted traffic don already get.
The CLI (command line interface) agree with this one. As of 0.1.0-rc.6, e deliberately no dey support --host 0.0.0.0 and e dey exit with usage error instead of starting, and the newer 0.2.2-alpha.2 still dey reject am for startup. That refusal na feature, so no go find patch wey go remove am.
Two other deployments dey reasonable if tunnel no work for you. Put the box for private overlay network so e go get address wey only your own devices fit route to, wey be wetin a self-hosted Headscale control server dey give you. Or put reverse proxy for front wey go authenticate the request before e reach port 3080, for example an Authentik single sign-on server wey dey do forward auth. Reverse proxy wey no get auth for front no be security control. Na just longer URL.
Access the web UI through an SSH tunnel
Run this for your laptop, no be for the server.
ssh -N -L 3080:127.0.0.1:3080 you@your-server-L dey open port 3080 for your laptop and forward anything wey connect to am through the encrypted SSH session. The 127.0.0.1:3080 part dey resolve for the server, so the connection go reach the harness from loopback, exactly as if you dey sit for the machine. -N mean say make e no start remote shell, because na the forward alone you want.
Then open http://127.0.0.1:3080 for your local browser. For release wey get launch token, open the complete startup URL wey the server print instead, including the token; when you start the harness through SSH, e still print the startup line but e no try open browser for you. If port 3080 don already dey busy for your laptop, change the number for the left side: ssh -N -L 3180:127.0.0.1:3080 you@your-server, then browse to http://127.0.0.1:3180, and change the port for tokenized URL the same way. The number for the left side na local one, while the number for the right side belong to the server, so na only the left one you go change.
Save am for ~/.ssh/config and stop typing am.
Host dsh
HostName 203.0.113.10
User deploy
IdentityFile ~/.ssh/id_ed25519
LocalForward 3080 127.0.0.1:3080After that, ssh -N dsh go start the tunnel. If browser report say connection dey refused, e usually mean say the tunnel dey up but nothing dey listen for the far side, because SSH dey forward the port whether the harness dey run or not. Check the server with the ss command wey dey above.
Make the harness continue to run after you log out
One npx command dey die once your shell close. One systemd user service go survive, and e go bring the harness back if e crash or if the system reboot. The unit wey dey here dey intentionally simple; if you want make the harness run under im own restricted account, with the version wey you lock inside the unit and logs wey you fit search well, the headless systemd setup for dsh explain everything for inside.
loginctl enable-linger $USER
mkdir -p ~/.config/systemd/user
command -v dshenable-linger dey important because user services normally dey stop when your last session end, so without am, the harness go die the moment you close the tunnel. Use the absolute path wey command -v dsh print and put am for the unit, because systemd no dey search the PATH wey your login shell dey build.
[Unit]
Description=DeepSeek Harness web UI
After=network-online.target
[Service]
Type=simple
WorkingDirectory=%h/projects/site
ExecStart=/usr/local/bin/dsh web
Restart=on-failure
RestartSec=5
[Install]
WantedBy=default.targetWorkingDirectory no be just for show. The dsh process dey use the directory wey you start am from as the default filesystem location, so if you start the service for wrong place, the agent go point go wrong folder. But that directory na just default: one fresh Web UI no get workspace wey dem select until you add one under Choose workspace, and the project guide tell you say make you add the directory wey you start dsh inside. The guide no get version marker, and the same wording dey inside 0.1.0-rc.7, so expect am for the pinned release too. Set the unit directory to the project, then select that workspace for the UI.
systemctl --user daemon-reload
systemctl --user enable --now dsh
systemctl --user status dshIf unit no gree start, na almost always because of wrong ExecStart path or Node version wey the binary no accept, and journalctl --user -u dsh -n 50 go tell you which one. For release wey get launch token, the journal na where the tokenized startup URL dey land, and every restart go print new one. The same pattern apply to keeping any coding agent alive on a VPS, and the ways wey e fit fail dey the same.
Plugin fit do wetin
A plugin na module wey dey add services, typed events, and effects wey fit reverse to shared context. Na these extension points you need read well:
- register model provider for
ctx.llm - add tools wey model fit use for
ctx.tools - provide shell backend behind
ctx.shell - provide filesystem access or policy behind
ctx.fs - register commands wey human fit run for
ctx.commands - run background work through
ctx.jobs - use
ctx.sandboxbackend wrap processes wey e spawn - intercept requests and tool calls through
agent/*andtools/*events - extend durable session state
- control UI through
ctx.agents
Read this list the way attacker go read am. Plugin fit provide filesystem layer and shell layer, and e fit dey inside every tool call wey model make. No permission dialog dey between plugin and these connection points, because plugin na ordinary Node code wey load for the same process with everything else. To install plugin mean say you dey run stranger code with your agent permissions, and your agent permissions na the permissions of your Unix user.
Na the same trust decision you dey make when you attach an MCP server to agent for VPS, where MCP mean model context protocol. Na also why to run coding agent safely for VPS dey start with the account wey e go run under, instead of starting with the model, and why npm supply chain attacks dey hit servers hard: the install step na where breach happen, and nothing go prompt you.
Where plugins come from
Plugins dey live inside profiles. Profile na one named composition wey dem store under $DSH_HOME, wey default to ~/.dsh, and every profile directory dey keep the out-of-tree plugins wey e install. The CLI dey manage dem by forward your arguments straight go pnpm, and e dey use the profile directory as the working directory.
dsh plugin --profile web add github:deepseek-harness/turtle-ui
dsh plugin --profile web remove turtle-uiBecause the arguments dey reach pnpm without any change, add, remove, update and why dey behave the same way dem dey behave for any pnpm project, and plugin fit be npm package or GitHub reference. pnpm must dey for the PATH first. For Node 24, corepack enable pnpm dey put am there. Node 25 stop to dey distribute Corepack, so for Node 25 and newer, use npm install -g pnpm to install am instead.
Discovery dey work through GitHub topic. Plugin authors dey add the dsh-plugin topic to their repository, and na by browsing that topic you go take find wetin dey available. Topic na label wey author dey apply to their own repository. Nobody dey review am and nobody dey sign am, and the topic page dey rank based on stars, wey just dey measure popularity, no be safety.
Current releases also get experimental Custom plugin indexes bundle. Switch am on under Plugins and one Discover tab go allow you add index, wey be JSON file wey dem fetch from URL wey you provide. The harness no dey ship any index of its own, e dey check the file format but e no dey check the titles, authors or sources wey market claim, and anybody wey dey the network path fit rewrite index wey dem serve over plain http:. Market na catalogue wey person write, e no be review.
Four habits go help you manage this one. Read the source before you install, because most plugins small reach make you read dem for ten minutes. Pin the exact version or commit instead of tracking branch. Run the harness under user wey no get any other thing, for VPS wey you ready to rebuild. Give the agent its own API key with its own spending limit, separate from the key wey your production services dey use. When you sit down to read plugin and you want know which files get risk, the walkthrough of vetting a dsh plugin go show you the manifest, the entry point and the extension points wey plugin dey register.
If you prefer to compare designs before you commit to one, the Omnigent multi-agent harness dey answer the same problem with different structure, and the trade-offs go clear once plugins enter play. If you come end up keeping two or three of dem for the same box instead of picking one winner, putting every harness behind one self-hosted API go save you from one tunnel per port, but you go get one extra service wey you must bind to loopback and give real password from day one.
Wetin fit spoil first
Node too old. The project main branch and 0.2.2-alpha.2 need Node 24.9 or newer. The CI tests Node 24, plus compatibility check for Node 24.9 and Node 26. E no get anything for Node 22. Older tagged releases up to 0.2.1-rc.4 accept Node 22.19, but the project no longer builds with am. Older runtime go fail for startup because the code dey use syntax and APIs wey e no get. Run node --version before you do anything else.
Port 3080 don already dey use. E fit be second harness, old process, or another application wey dey use 3080 too. Find am with ss -tlnp | grep 3080, then stop am or start the harness for another place with dsh web --port 3180. --port belong to the web app, so e go come after web.
Browser no fit connect through tunnel. Confirm say you browse go 127.0.0.1, no be the server public address, because the forwarded port dey only for your laptop. Then confirm say the harness dey listen for the server, because SSH go set up the forward whether anything dey answer for the far end or not.
dsh plugin fail immediately. The command na wrapper around pnpm, so if pnpm binary no dey, e go stop before any plugin work start. For Node 25 and newer, na the default state be that, because Corepack no longer ships with Node.
Agent no fit see your project. Fresh Web UI no get workspace selected until you add one. The session composer no go available until you do that. The process dey use the directory wey e start from as the default filesystem location only. So if unit WorkingDirectory dey your home directory, agent go point to your home directory. Choose the workspace for the UI, or correct the unit and reload am.
FAQ
E safe make I expose DeepSeek Harness web UI for port 3080?
No. For the pinned 0.1.0-rc.6, the web server no get im own login system, and the agent wey dey behind am fit edit files and run shell commands as the user wey start the process, plus your provider API key dey stored for the same disk. From 0.2.0-rc.2 go front, including 0.2.2-rc.1, npm latest release wey drop for 12 November 2026, and the newer 0.2.2-rc.2 prerelease, the startup URL wey dem print get launch token wey browser dey trade for session cookie. But that URL na credential wey plain HTTP dey send without encryption, so the advice no change. Keep the listener for 127.0.0.1 and use SSH tunnel reach am. Private overlay network, or reverse proxy wey dey authenticate every request before e reach the port, sef go work. As of version 0.1.0-rc.6, the CLI dey refuse --host 0.0.0.0 and e dey exit with usage error, and 0.2.2-alpha.2 still dey reject am, wey show you wetin the authors think about that idea.
I need DeepSeek API key, or I fit use local model?
Any one go work, because the harness na runtime, no be model. Under Settings come go Models, you fit paste key for the DeepSeek card or catalog provider, or choose "Add model provider" then "Custom model API" and give am base URL wey fit talk OpenAI Chat Completions, OpenAI Responses, Anthropic Messages or Gemini API. Local Ollama server dey answer for http://127.0.0.1:11434/v1/ and e dey accept any string for the API key field. Keys dey land for $DSH_HOME/.credentials.yaml, wey dey default to ~/.dsh/.credentials.yaml.
Wetin installing DeepSeek Harness plugin actually give the plugin?
The permissions of the account wey dey run the harness. Plugin na Node code wey dem load into the same process, and the extension points include shell backend, filesystem layer, tool registry and the events wey dey wrap every tool call. Nothing dey sandbox plugin comot from those areas unless the plugin provide the sandbox by imself. Read the source code before you install, and run the harness as user wey no get access to anything wey you value.
Which version I suppose install, and e go continue to work?
Install exact version, for example npx @deepseek-ai/dsh@0.1.0-rc.6 web. That one na the August 2026 version wey dem pin here. For 12 November 2026, npm latest tag dey point to 0.2.2-rc.1; GitHub list the newer prerelease 0.2.2-rc.2, wey drop for 11 November, and the current CLI require Node 24.9 or newer. The project call imself developer preview and dem talk say breaking changes go happen, so command wey you no pin fit behave anyhow from one day to another. Check the repository before you upgrade, and expect say config keys and plugin interfaces fit move as long as the version still dey start with 0.