Self-hosted AI agents, deployed in one click

From checkout to a working agent: what actually happens

This is the step most one-click pages leave out: the moment after you have picked an image and the server exists, and you are looking at it wondering what to do next. Here is the complete path, with nothing assumed.

1. At checkout, choose your image (OpenClaw, Hermes, or one of the supporting stacks), your plan, and your region, then either upload an SSH public key or set a root password, whichever you prefer to log in with.
2. Provisioning takes about ninety seconds. When it finishes, your client area and your email both receive the server's IP address, the image that was installed, and confirmation of whichever login method you chose.
3. Log in over SSH (with your key, or the password you set) using any SSH client, including the browser-based terminal in your client area if you do not have one installed locally. This is the same login you would use for any KryoHost VPS, nothing agent-specific about it.
4. Open a tunnel to the admin interface. Because the admin UI is bound to localhost for security, reaching it from your own browser means forwarding that port over SSH, one command, documented with the exact syntax in the welcome email and the knowledgebase article for your image.
5. Supply the image's specific requirement: for OpenClaw, that is a model provider's API key entered into its config; for Hermes, that is running the single documented command to pull the model weights you have chosen.
6. Start using it. From here it behaves like the documentation for that specific image describes, OpenClaw takes tasks through its interface, Hermes answers requests on its API endpoint.

Available agent images

Each image is maintained against upstream releases and rebuilt when a security update lands. The image version and its upstream commit are recorded in /etc/kryohost-image so you always know what you are running.

All images ship with unattended security updates enabled, SSH key authentication only, and a firewall denying inbound traffic by default. What they do not ship with is an opinion about your application architecture, they are starting points.

The infrastructure behind KryoHost 1-Click AI Agents

Every conversation about self-hosted AI agent hosting eventually comes back to hardware, and for good reason. You can tune a stack endlessly, but you cannot compensate in software for a disk that is queueing, a CPU that is oversubscribed four times over, or an upstream carrier that routes your visitors halfway around the planet before delivering the first byte. KryoHost builds from the metal upward precisely because the metal sets the ceiling on everything above it.

Our our agent images fleet runs on dual-socket AMD EPYC and Intel Xeon Scalable nodes with ECC memory, paired exclusively with enterprise NVMe drives in RAID-10. There is no SATA tier hiding behind the marketing copy, and no "NVMe-accelerated" wording that quietly means a small cache in front of spinning disks. NVMe changes the character of a server rather than simply making it faster: random read latency drops from milliseconds to microseconds, so the database queries that dominate page-generation time on self-hosted AI agents, inference endpoints and automation stop being the bottleneck. In our own benchmarks, moving an unchanged WordPress install from a SATA SSD node to an NVMe node cut median time-to-first-byte by 38% without a single line of code being touched.

Capacity planning is the unglamorous half of the story. We cap node density well below what the hardware could theoretically carry and we alarm on sustained CPU steal, disk queue depth and memory pressure rather than waiting for customers to open tickets. When a node crosses its threshold, new provisioning stops on that node and workloads are rebalanced. That is why the phrase "the server was fine, your site is just heavy" is one you will not hear from our team, if steal time is climbing, that is our problem to solve, not yours to absorb.

Network, peering and routing

KryoHost operates across Tier-III and Tier-IV facilities in 47 customer-selectable countries. Each core site is multi-homed across at least three Tier-1 carriers and connected to the dominant regional internet exchange, so traffic reaches your visitors through the shortest sensible path rather than the cheapest available one. Blended transit is convenient for a provider and mediocre for a customer; direct peering costs more and is the reason a visitor in Frankfurt does not have their packets tour Europe before they see your homepage.

• Redundant power: N+1 UPS with diesel generators tested under load monthly, not merely started and logged.
• Redundant cooling: hot/cold aisle containment with N+1 CRAC units and independent chilled-water loops.
• Redundant network: no single upstream carries more than 40% of a site's traffic, so losing one is a re-route, not an outage.
• Physical security: biometric access control, mantraps, 24/7 on-site security and 90-day CCTV retention.
• Independent audit: facilities hold SOC 2 Type II and ISO 27001 certification, with reports available to enterprise customers under NDA.

The practical effect of all this is measurable rather than decorative. Our published 99.99% uptime SLA is backed by service credits, and our public status page records every incident, including the ones that lasted four minutes and that nobody noticed. A provider that only publishes its good quarters is not publishing anything useful.

Security, backups and disaster recovery

Most sites are not compromised by a determined attacker studying them for weeks. They are compromised by automated scanners walking the entire IPv4 space, trying known plugin vulnerabilities and credential lists against everything that answers on port 443. Defence against that reality is layered and mostly boring, which is exactly why it works, and why it should already be switched on when your deployment is delivered rather than sold as an upsell after the incident.

What runs by default

• Web application firewall: ModSecurity with continuously updated OWASP and vendor rule sets, tuned per stack to keep false positives low.
• Malware scanning: Imunify360 scans on write and on schedule, quarantines known-bad files and can auto-clean common injections.
• Brute-force protection: progressive rate limiting and automatic IP throttling on login endpoints, SSH and mail services.
• DDoS mitigation: always-on L3/L4 scrubbing at the network edge, with L7 protection available on request.
• Account isolation: CageFS confines each account to its own virtualised filesystem, so one compromised neighbour cannot read another's data.
• Free TLS: automatically issued and renewed certificates with modern cipher suites and HSTS available in one click.
• Patched kernels: KernelCare applies security patches without reboots, removing the trade-off between staying current and staying up.

Backups, and what they actually cover

Website files on every deployment are backed up twice a week, with email accounts backed up on a separate schedule so both are protected without either job blocking the other. You can also generate and download a full account backup at any time from cPanel, and our support team can help with a restore if you need one.

A twice-weekly schedule is not the same as continuous protection, and we would rather you knew that than assumed otherwise. If you publish or sell something several times a day, keep your own more frequent backup for the hours between our scheduled runs, particularly for a database that changes constantly, such as an order table.

Compliance-wise, our infrastructure supports GDPR data-residency requirements through EU-only regions, and we sign Data Processing Agreements on request. For workloads touching payment data, our environment is PCI-DSS ready, you still own the compliance of your own application, but the platform underneath will not be the reason an assessment fails.

Performance, Core Web Vitals and what actually moves rankings

Search engines have been explicit that page experience is a ranking input, and Core Web Vitals are the measurable part of that signal. What gets lost in most self-hosted AI agent hosting marketing is which of those metrics hosting can genuinely influence. Being precise about it saves you money, because it tells you when to buy a bigger deployment and when to fix your front end instead.

Read that table honestly and a useful rule emerges: hosting owns the first 200–400 milliseconds and your front end owns most of what follows. That is not an argument for cheap hosting (those first milliseconds are a floor that nothing downstream can undo) but it is an argument against believing a plan upgrade will rescue a page carrying eleven tracking scripts and a 4 MB uncompressed hero image.

The caching layers we ship enabled

KryoHost our agent images ships with a caching stack that is configured on day one rather than left as an exercise. Requests are answered at the shallowest layer that can serve them correctly, and each layer that handles a request is one your origin never sees.

1. Edge CDN: 120+ points of presence serve static assets from the PoP nearest your visitor, typically within 10–30 ms.
2. Full-page cache: LiteSpeed Cache stores rendered HTML in RAM, so repeat visits skip PHP and the database entirely.
3. Object cache: Redis keeps expensive query results and transients in memory instead of round-tripping to MySQL.
4. OPcache: compiled PHP bytecode is held in memory, removing parse and compile cost from every request.
5. Browser cache: long-lived immutable headers on fingerprinted assets mean returning visitors re-download almost nothing.

Compression and protocol choices are handled at the edge too. Brotli is preferred over gzip where the client supports it, HTTP/3 with QUIC is enabled by default so lossy mobile networks stop paying the TCP head-of-line penalty, and TLS 1.3 removes a full round trip from the handshake. None of these require a support ticket; they are the default configuration on every deployment.

Support that answers the question you actually asked

Support is where hosting companies quietly differ most, and where the difference is hardest to evaluate before you buy. Every provider claims 24/7 availability. Far fewer will tell you who is on the other end at 03:00 on a Sunday, how many tickets that person is holding, or what percentage get resolved without being escalated into a queue you cannot see.

KryoHost staffs Linux system administrators across three timezones. There is no offshore first line whose job is to send you a knowledgebase link and close the ticket. Our commitment is a first response inside < 15 min on anything service-affecting, escalating automatically to on-call if that is missed. When something does need a second pair of hands it moves to a named senior engineer and you are told who owns it, not dropped into a silent queue.

What is included, and what is genuinely out of scope

Being clear about scope up front prevents the most common support frustration: discovering after an incident that the thing you assumed was covered never was.

Support reaches you through live chat, tickets and email, and (on business and VPS plans) a scheduled screen-share call when something genuinely needs a live conversation. The same standard of engineer answers at any hour, which matters more than it sounds when you are describing an intermittent fault under pressure at two in the morning.

How to choose the right 1-Click AI Agents plan

Buying hosting badly usually happens in one of two directions. Either you buy the cheapest thing available, hit its ceiling within a quarter and migrate under pressure; or you buy three tiers above what you need and pay for headroom that sits idle for years. Both are avoidable with about ten minutes of honest arithmetic.

Start from traffic, not from storage

Storage is the number hosting companies advertise because it is cheap to give away and easy to compare. It is almost never the constraint. A well-built content site with a few hundred posts occupies under 2 GB. What actually determines whether you outgrow a deployment is concurrency, how many requests need generating simultaneously at your busiest minute, not your busiest month.

Take your peak day, look at the busiest hour, and assume the busiest minute carries roughly three times the average minute in that hour. If that number is under about 60 dynamic requests, entry our agent images is genuinely sufficient. Between 60 and 300, you want guaranteed CPU. Above that, you want a VPS or a container-isolated business plan where noisy neighbours cannot reach you at all.

Read the renewal price before you read the headline price

The introductory rate on a self-hosted AI agent hosting plan is a customer-acquisition cost, not a price. What you will actually pay is the renewal figure, multiplied by however many years you keep the site. We publish both on every plan card on this page for exactly that reason. A $19.99 entry price that renews at four times the rate is not cheap hosting; it is a loan against your future attention.

Compare on total cost over three years, including domain renewal, SSL if it is not free, backups if they are billed separately, and any migration fee at the point you eventually leave. That last one matters: a host that charges to let you go has told you something about its confidence in the product.

1-Click AI Agents terminology, decoded

Hosting vocabulary is unusually good at making simple ideas sound complicated, and occasionally at making limited products sound generous. Here is what the terms on a self-hosted AI agent hosting comparison page actually mean in practice.

If a provider will not define its soft limits in writing, treat that as the answer. Ours are published in the Terms of Use, in plain language, with the numbers included.

Billing, guarantees and cancellation

Billing should be the least interesting part of your relationship with a hosting provider. It becomes interesting only when a provider uses it as a retention mechanism, automatic multi-year renewals, cancellation flows hidden behind three menus, refunds that require a phone call during business hours in a timezone you do not live in. We have deliberately built ours to be dull.

• Cancel from the panel: one button in your client area. No ticket, no phone call, no "let me transfer you to retentions".
• Pro-rata credit on upgrades: move up mid-cycle and the unused portion of your current term is credited automatically.
• Renewal reminders: 30, 14 and 3 days before any charge, with the exact amount stated in the email.
• No hidden setup fees: the price on the plan card is the price at checkout, before tax.
• Free data export: a full account archive on request at any time, including after cancellation, for 30 days.

We accept cards, PayPal, bank transfer and major cryptocurrencies. Invoices are issued with full tax details for business accounts, and annual billing carries a genuine discount rather than a discount against an inflated monthly rate.

On the SLA: 99.99% availability is measured monthly at the network edge, excluding scheduled maintenance announced at least 72 hours in advance. Fall below it and you receive service credits automatically, you do not have to notice, calculate and claim. An SLA that requires the customer to police it is not much of a commitment.