Last updated: September 22, 2026
Here are the best DeepSeek Harness hosting providers in 2026 by use case. Hostinger is the easiest option, with a one-click dsh template. Hetzner Cloud has the lowest self-hosted bill. Railway works if you want a password-protected container without managing Linux. DigitalOcean is the best choice if you learn by following tutorials. All four keep the harness running after your laptop goes to sleep.
That covers most readers. The rest of this guide explains three things: why hosting dsh takes more than one npx command, what each provider costs once any promo price ends, and which setup fits your skill level.
We checked every price and technical requirement in this guide against provider pricing pages, official documentation, and the DeepSeek Harness repository in September 2026. Setup times are estimates for someone who is comfortable in a terminal. We did not run benchmarks, so none are presented here.
Quick picks: best DeepSeek Harness hosting providers at a glance
| Provider | Type | Starting price | Setup effort | Best for |
|---|---|---|---|---|
| Hostinger VPS | VPS with dsh template | $6.49/mo on a 24-month term (renews at $11.99) | About 10 minutes | Beginners who still want root access |
| Hetzner Cloud | Self-managed VPS | €5.49/mo (CX23) | 30 to 60 minutes | Lowest steady monthly cost |
| Railway | Container platform | Usage-based | About 10 minutes | Always-on dsh without managing a server |
| DigitalOcean | Self-managed VPS | $4/mo (about $12 for 2 GB) | 30 to 60 minutes | Tutorials and free starter credit |
| Vultr | Self-managed VPS | $5/mo | 30 to 60 minutes | Choosing a region near your team |
| Contabo | Self-managed VPS | About €4.50/mo for 8 GB RAM | 30 to 60 minutes | Most RAM per dollar |
| Oracle Cloud Free Tier | Free ARM VM | $0 | 45 to 90 minutes | Learning on zero budget |
What is DeepSeek Harness?
DeepSeek Harness (dsh) is DeepSeek’s open-source agent harness. It is the runtime that turns a language model into an agent that can edit files, run shell commands, search the web, and keep working through long tasks. DeepSeek released it as a developer preview on August 13, 2026, alongside its V4 models, under the MIT license.
The design idea is that everything is a plugin. dsh is built on the Cordis kernel. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and even the UI are all swappable plugins that you configure without forking the code. It can also hand work to Claude Code or Codex as subagents, so it can sit above those tools as well as compete with them.
The project grew quickly. The official DeepSeek Harness repository passed 230,000 GitHub stars within about six weeks. If you are comparing dsh with other agent runtimes, our guide to the best AI agent hosting providers covers the wider market.
Why DeepSeek Harness needs proper hosting
Running npx @deepseek-ai/dsh web on a laptop works fine for a demo. Several parts of the design only pay off on a machine that never sleeps.
The Web UI listens on 127.0.0.1:3080 and has no built-in login. The CLI refuses to bind to a public address because doing so would expose remote code execution to the network. That default is correct. It also means you need a reverse proxy with TLS and a password in front of dsh before you can reach it from another device.
Scheduled loops only fire while the process is running. The append-only session log, which powers resume, fork, and replay, only keeps its value on disk that survives restarts.
The project is also changing quickly. The README warns that breaking changes will happen, and every npm release so far has been a release candidate. Good hosting for dsh makes it easy to pin a version, roll back, and rebuild.
What a DeepSeek Harness server needs
| Requirement | Why it matters |
|---|---|
| Node.js 22.19+ or 24+ | The package declares these engine versions. Node 20 can hang during install without a clear error. |
| Persistent disk | Sessions, settings, and keys live in the dsh data directory (~/.dsh by default) |
| Reverse proxy with TLS and auth | The UI is loopback-only and has no login of its own |
| Process supervisor | systemd or a Docker restart policy brings dsh back after a crash or reboot |
| Outbound HTTPS | The harness calls hosted model APIs such as DeepSeek, Anthropic, OpenAI, or OpenRouter |
| Backups | Losing the session log means losing your replay history |
| GPU | Not needed unless you also run a local model on the same server |
How much RAM does DeepSeek Harness need?
The harness itself is light. One community template measured its NixOS build at about 135 MiB while idle. Installing packages at runtime, running sandboxed tools, and keeping several sessions open push memory use up quickly, though.
| Workload | Suggested size |
|---|---|
| Testing over an SSH tunnel | 1 vCPU, 1 GB RAM |
| One always-on agent, light use | 1 to 2 vCPU, 2 GB RAM |
| Daily use with sandboxed tools | 2 vCPU, 4 GB RAM |
| Several sessions or heavy tool runs | 4 vCPU, 8 GB RAM |
| A local model on the same box | 16 GB RAM or more (the model sets the floor) |
How we picked the best DeepSeek Harness hosting providers
We scored each host on six points:
- How much of the proxy, TLS, and authentication work the host does for you
- Time from sign-up to a working, remotely reachable agent
- Price after any promotional term ends, not just the banner price
- Support for persistent storage and backups
- Whether current Node.js versions and root or container access are available
- How well the host fits a developer preview that ships breaking changes
This is the same approach we use in our guide to the best VPS hosting providers. We compare published specs and renewal pricing, and we treat vague claims with caution.
Comparison table: security and maintenance work
| Provider | Proxy and TLS done for you | Login in front of dsh | Backups | Price changes at renewal |
|---|---|---|---|---|
| Hostinger VPS | Yes, via template | Yes, via template | Included snapshots and backups | Yes |
| Hetzner Cloud | No | No | Paid add-on | No promo pricing |
| Railway | Yes, via template | Yes, basic auth | Volume-based | Usage-based |
| DigitalOcean | No | No | Paid add-on | No promo pricing |
| Vultr | No | No | Paid add-on | No promo pricing |
| Contabo | No | No | Manual or paid | No promo pricing |
| Oracle Free Tier | No | No | Manual | Free |
1. Hostinger VPS: best overall DeepSeek Harness hosting
Hostinger added a DeepSeek Harness template to hPanel. You select it under application templates on your VPS and click Install. hPanel then sets up the web interface, the reverse proxy, and the TLS certificate. You sign in with admin as the username and your VPS password, as described in Hostinger’s DeepSeek Harness setup guide.
That is why Hostinger takes the top spot. It is the only VPS in this list that handles the riskiest part of running dsh, which is exposing it safely, and it still gives you root access for plugins, cron jobs, or Docker containers.
Pricing is the catch. KVM 1 (1 vCPU, 4 GB RAM, 50 GB NVMe) costs $6.49 per month on a 24-month term and renews at $11.99. KVM 2 (2 vCPU, 8 GB RAM, 100 GB NVMe) costs $8.99 and renews at $14.99. Plan your budget around the renewal price.
Pros:
- A one-click template handles proxy, TLS, and login
- Plenty of RAM for the price, with NVMe storage
- Root access for plugins and custom tooling
Cons:
- Big renewal jumps, and the lowest rates require long prepaid terms
- The agent runs with full server permissions, so your VPS password protects everything
- Some users report CPU throttling during sustained heavy load
Best for developers who want a working dsh URL in minutes but still want a real server. Our full Hostinger review covers the renewal math in detail.
2. Hetzner Cloud: cheapest self-hosted DeepSeek Harness VPS
Hetzner is still the price leader for do-it-yourself hosting, even after two price increases in 2026. The CX23 plan has 2 vCPU, 4 GB RAM, 40 GB NVMe, and 20 TB of EU traffic for €5.49 per month (excluding VAT) at the rate that applied from June 15, 2026. That is enough to run dsh with sandboxed tools and some room to spare.
You get a clean Linux server and nothing else. There is no dsh template and no guided setup. You install Node.js 22.19 or newer, pin a dsh version, write a systemd unit, set up Caddy or nginx with a password, and lock down the firewall yourself.
The CX line is available in EU regions only. US and Singapore regions use the more expensive CPX plans, which come with much smaller traffic allowances.
Pros:
- Lowest steady monthly cost for a production-grade 4 GB server
- No gap between promo and renewal pricing
- Free cloud firewall and hourly billing while you test
Cons:
- Everything is manual, from proxy to backups
- The cheapest plans are EU-only
- Prices have risen twice this year, so check the current rate before ordering
Best for experienced Linux users who want the smallest bill and don’t mind a setup evening.
3. Railway: best container hosting for DeepSeek Harness
Railway has community templates that deploy dsh in a container with a persistent volume, in Ubuntu and NixOS versions. The template design is careful. dsh stays bound to loopback, and a Caddy proxy handles the public port and enforces HTTP basic auth. If you leave the password blank, the container generates a strong one on first boot and prints it once to your deploy log. You cannot switch the password gate off.
The NixOS version lets the agent install nixpkgs packages at runtime without root. It idles at about 135 MiB, but the template author recommends at least 1 GB of memory because package installs cause memory spikes.
Billing is usage-based, so costs depend on how much CPU and memory the container uses. Keep in mind that these templates are community-built. They are not official releases from Railway or DeepSeek. Check which dsh image version the template pins before you rely on it.
Pros:
- Secure access with a password gate, with no Linux administration
- Persistent volume keeps sessions across redeploys
- Fast to launch and easy to redeploy
Cons:
- Community-maintained templates that may lag upstream releases
- Usage billing is harder to predict than a flat VPS price
- Less control than a full server
Best for developers who want an always-on dsh URL and would rather manage containers than servers. See the Railway DeepSeek Harness template.
4. DigitalOcean: best for documentation and starter credit
DigitalOcean’s Basic Droplets start at $4 per month. For an always-on agent, 2 GB of RAM is a sensible minimum, which puts you at about $12 per month. Since January 2026, Droplets have been billed per second with a monthly cap, which suits short test servers.
DigitalOcean’s biggest strength is its tutorial library. If you have never configured Caddy, systemd, or a cloud firewall, you can find a detailed guide for each step. New accounts typically get $200 in credit for 60 days, which covers a long trial.
Pros:
- Excellent tutorials for every self-hosting step
- Free cloud firewalls and simple networking
- Starter credit covers weeks of testing
Cons:
- Less RAM per dollar than Hetzner or Contabo
- Backups cost extra
- No dsh template, so you secure it yourself
Best for developers who learn by following guides and want a mainstream cloud.
5. Vultr: best for choosing a region near your team
Vultr runs 32 cloud regions, which is useful when your team or your model provider is far from Europe. Placing dsh close to its users makes the Web UI feel faster and cuts round-trip time on webhooks.
Plans start at $5 per month for 1 GB of RAM. That is fine for testing, but choose 2 GB or more for daily use. Billing is hourly, so you can try dsh in several regions and delete the servers afterward.
Pros:
- Wide regional coverage
- Hourly billing for experiments
- Reserved IPs for stable proxy addresses
Cons:
- The 1 GB entry plan is tight for sandboxed tools
- Fewer tutorials than DigitalOcean
- Everything security-related is on you
Best for distributed teams that care more about latency than the lowest price.
6. Contabo: most RAM per dollar
Contabo’s Cloud VPS 10 has 4 vCPU cores, 8 GB RAM, and 75 GB NVMe for about €4.50 per month, based on early 2026 pricing. Check the current rate for your region before ordering. That much memory is useful if you run many concurrent dsh sessions, heavy sandboxed tools, or a small quantized model next to the harness.
The trade-offs are well known. The network port is limited to 200 Mbit/s on the entry plan, performance on shared CPUs can vary, and users often report slower provisioning and support than on premium clouds.
Pros:
- Much more RAM than competitors at a similar price
- Large storage allocations
- EU and US regions
Cons:
- Performance and support are less consistent
- No dsh tooling or templates
- You need to set up your own offsite backups
Best for power users who value memory headroom more than polish.
7. Oracle Cloud Free Tier: best free option for learning
Oracle’s Always Free tier includes Ampere A1 ARM compute with up to 4 OCPUs and 24 GB of RAM at no cost. dsh runs on Node.js, and Node 22 works well on ARM64, so you can run a full personal agent without paying anything.
You pay with effort instead. Card verification often fails during sign-up, and popular regions run out of free capacity. Idle instances can be reclaimed. The console is built for enterprises, and Oracle has two firewall layers (the security list and the OS firewall), which catches out many first-time users.
Pros:
- $0, with more RAM than any paid entry plan here
- Real infrastructure, not a time-limited trial
- Plenty of room for experiments
Cons:
- Difficult sign-up and limited capacity
- Idle reclamation risk
- Two firewall layers to configure
Best for students and hobbyists. If you enjoy this kind of setup, our list of the best self-hosted open source applications has more projects to run on the same box.
What DeepSeek Harness hosting really costs
Promo prices make hosts look cheaper than they are. This table shows hosting costs only.
| Provider and plan | Price now | Price after promo | Yearly cost at regular price |
|---|---|---|---|
| Hostinger KVM 1 | $6.49/mo (24-month prepay) | $11.99/mo | About $144 |
| Hostinger KVM 2 | $8.99/mo (24-month prepay) | $14.99/mo | About $180 |
| Hetzner CX23 | €5.49/mo | Same | About €66 |
| DigitalOcean 2 GB | About $12/mo | Same | About $144 |
| Vultr 1 GB | $5/mo | Same | About $60 |
| Contabo Cloud VPS 10 | About €4.50/mo | Same | About €54 |
| Oracle Free Tier | $0 | $0 | $0 |
| Railway | Usage-based | Usage-based | Varies |
Model API tokens are billed separately by your model provider. VentureBeat reported that DeepSeek moved to peak and off-peak API pricing in mid-August 2026, so running scheduled loops in off-peak hours can lower that bill.
Your own time is the cost most comparisons leave out. At a $40 hourly rate, two hours spent securing a bare server costs more than a year of the price difference between Hetzner and Hostinger’s template-based setup.
Which DeepSeek Harness host should you choose?
| If you are… | Choose | Why |
|---|---|---|
| A beginner who wants root access | Hostinger VPS | The template handles the risky exposure step |
| Comfortable with Linux and watching costs | Hetzner Cloud | Lowest steady price for 4 GB |
| Happier with containers than servers | Railway | Password gate and persistent volume by default |
| Learning by following tutorials | DigitalOcean | A guide exists for every step |
| Serving a global team | Vultr | 32 regions |
| Running many sessions at once | Contabo | 8 GB RAM at an entry-level price |
| Working with no budget | Oracle Free Tier | 24 GB RAM for $0 |
If you would rather hand off server work entirely, our ranking of the best managed hosting provider options explains what fully managed plans include and what they cost.
How to self-host DeepSeek Harness safely
If you choose a bare VPS, follow this order:
- Install Ubuntu 24.04 and Node.js 22.19 or newer. Node 20 is not supported.
- Pin an exact dsh version instead of the moving
latesttag, for examplenpx @deepseek-ai/dsh@<version> web --no-open. Check npm for the current release first. - Keep dsh bound to
127.0.0.1. For quick tests, reach it through an SSH tunnel:ssh -L 3080:127.0.0.1:3080 user@your-server. - For access from any browser, put Caddy or nginx in front of dsh with TLS and password authentication. dsh checks the Host and Origin headers, so your proxy has to rewrite them or you have to register your domain with the harness’s trusted-host option. Flag names change between releases, so check
--helpfor your pinned version. - Run dsh under systemd or Docker with an automatic restart policy, and keep the data directory (
DSH_HOME, which defaults to~/.dsh) on persistent storage. - Allow only ports 22, 80, and 443 through the firewall.
- Back up the data directory on a schedule.
The most common mistake is binding dsh to 0.0.0.0 to “make it work.” That puts an agent with a shell on the open internet. The second most common is running it in Docker without a mounted volume, which erases session history on every rebuild.
Frequently asked questions
What are the best DeepSeek Harness hosting providers for beginners?
Hostinger and Railway are the most beginner-friendly. Hostinger’s one-click template sets up the reverse proxy, TLS certificate, and login, and still gives you root access. Railway’s community templates keep dsh behind a password gate with a persistent volume. Both avoid the riskiest step, which is exposing dsh to the internet safely.
Do I need a GPU to host DeepSeek Harness?
No. By default, DeepSeek Harness calls hosted model APIs such as DeepSeek, Anthropic, OpenAI, or OpenRouter, so the server only runs agent logic, sessions, and tools. You need a GPU only if you run a local model on the same server with a tool like Ollama or vLLM.
How much RAM does DeepSeek Harness need?
The harness idles at a few hundred megabytes, but 2 GB is a sensible minimum for an always-on server. Choose 4 GB if you use sandboxed tools every day, and 8 GB if you run several sessions at once. A local model on the same machine needs much more.
Which Node.js version does DeepSeek Harness require?
DeepSeek Harness needs Node.js 22.19 or newer on the 22 line, or Node.js 24 and later. Users report that Node 20 hangs during installation without a clear error, so check your version before running the install command.
Why can’t I open the DeepSeek Harness Web UI on my VPS IP?
The Web UI only listens on 127.0.0.1:3080 by design, and it has no login of its own. Use an SSH tunnel for testing, or put a reverse proxy with TLS and password authentication in front of it. Do not bind dsh to a public address.
Is DeepSeek Harness ready for production?
Not fully. DeepSeek calls it a developer preview and warns that breaking changes will happen. Every npm release so far has been a release candidate. Pin a version, keep your configuration under version control, and test upgrades before you rely on them for critical work.
Can I move DeepSeek Harness to another host later?
Yes. Your sessions, settings, and plugin state live in the dsh data directory, which is ~/.dsh by default. Archive that directory, copy it to the new server, install the same pinned dsh version, and start the service again. This makes switching providers low risk.
If you run other agents alongside dsh, our comparison of the best Hermes Agent hosting providers uses the same criteria.
Final verdict on the best DeepSeek Harness hosting providers
Choosing among the best DeepSeek Harness hosting providers comes down to how much server work you want to do while the project is still changing every week.
If you want the least setup with full root access, Hostinger’s one-click template is the fastest safe path. If you enjoy Linux and want the lowest bill, Hetzner’s CX23 is hard to beat, and Oracle’s free tier costs nothing if you can get through sign-up. Railway suits people who prefer containers, and DigitalOcean, Vultr, and Contabo each fit a specific need.
Whichever you choose, pin your dsh version, keep the UI behind authentication, and back up ~/.dsh. For more guides like this, browse our AI and AI agents coverage.
0 Comments