Hostman Monthly Log: October 2026
Hi from the Hostman team, and welcome to the October issue of the Hostman Monthly Log.
October was mostly about AI and Kubernetes. Your agents can now remember the people they talk to, and a failed model no longer has to mean a failed request. On the Kubernetes side, you can keep a cluster's control plane off the internet entirely.
- Memory for AI Agents
- Fallback Models and Cheaper Cached Tokens
- Connect Your Cluster to AI Gateway
- Private Control Plane for Kubernetes
- 12 New Frameworks in App Platform
- More Updates
- Also Shipped
Memory for AI Agents Copy link
Agents used to start every new chat from scratch. Now they have long-term memory: they save what matters about each user, like their name, preferences, and the projects they're working on, and bring it into later chats.
The agent decides what's worth remembering, or you can ask it directly: "Remember that I write in TypeScript." Then open a new chat, ask "What language do I write in?", and it already knows. Facts that go out of date get updated, and anything you ask the agent to forget is deleted.
Each user has their own set of facts and only sees theirs. On the agent's Memory tab, you can review everything it has saved and add, edit, or delete facts. The limit is 1,000 facts per agent, shared across all its users.
Memory is off by default. Turn on Use memory when you create an agent, or click Enable Memory on the Memory tab of an existing one. It works in the agent's external chat and through the API, but not in the website widget. With memory on, agents can use more tokens.
Fallback Models and Cheaper Cached Tokens Copy link
When the model behind an AI Gateway request fails, your app gets an error. Fallback models give the request somewhere else to go. You can attach up to two fallback models to an API key. If the main model is down, times out, hits a rate limit, or can't fit the request into its context window, the gateway tries the first fallback, then the second.
You're billed at the rate of the model that answers, and failed attempts are free. The model field in the response shows which model answered. You'll find fallbacks on the key's Management tab, in the Fallback Models row.
Cached tokens now cost less. Long chats and agents with big system prompts resend a lot of the same text: the system prompt, the chat history, attached documents. When part of a request repeats, the model reads it from the cache, and those tokens cost about a tenth of the regular input price. This applies to both AI Gateway and AI Agents.
Connect Your Cluster to AI Gateway Copy link
Apps in your Kubernetes cluster that call AI models need an AI Gateway key. Until now, getting it there was a manual job: create the key, copy it, and add it as a Secret to each namespace. The new AI Connector does that for you.
Open your cluster's Management tab, find AI Connector, and create a key. Pick the namespaces your apps run in and how long the key should stay valid, and the connector adds it as a Secret to each of them. Your apps read the key from an environment variable or a mounted file.
When you regenerate the key, the Secrets keep their names, so your manifests stay the same. Delete the key, and its Secrets disappear from every namespace.
Private Control Plane for Kubernetes Copy link
Kubernetes clusters can now run without a public IP on the control plane. Turn off Public IP when you create a cluster, and its API can only be reached from the cluster's private network.
That's one less way in for attackers. Many security policies require it anyway: the part of your infrastructure that controls everything else has to be isolated. And if only services and people on the same private network use the cluster, it never needed a public IP in the first place.
You manage the cluster from inside its private network. To do that from your laptop, use a bastion host: a cloud server with a public IP in the same private network and region as the cluster. Connect to it over SSH and work with the cluster from there.
12 New Frameworks in App Platform Copy link
App Platform now supports 12 more frameworks, six for the backend and six for the frontend.
For backend apps: Echo and Fiber, aiohttp and Tornado, Quarkus and Micronaut.
For frontend apps, App Platform adds Astro, TanStack Start, Solid, SolidStart, Remix, and SvelteKit. Most of them also support server-side rendering, which you can turn on when you create the app.
More Updates Copy link
Stop and roll back deploys in staging environments
Staging environments (the Stands tab) now have the same deploy selector as their parent app. You can stop a running deploy or roll back to any commit you've deployed before.
Restart the runtime environment, not just the container
When a backend app stops responding, click Reboot in the top right corner and choose what to restart. Most of the time, a container restart does the job. If the app ran out of resources and that didn't help, restart the runtime environment, which reboots the server underneath too.
Python 3.15, Node.js 26, and more in App Platform
Your apps can run on Go 1.26 and 1.27, Node.js 26, Python 3.15, Java 27, Elixir 1.20, Bun 1.4, and Rust 1.99. Pick the version when you create an app.
More video, image, and text models in AI Gateway
New in the catalog this month:
- Video: Seedance 2.0 and 2.5, Kling 3.0, Kling O3, Kling 2.5 Turbo Pro, Veo 3.1, Hailuo 3, and Wan 3. Most come in several versions, and both Kling 3.0 and Kling O3 have a 4K one. Video generation works through the API with an AI Gateway key.
- Images: Seedream 5 Pro and Seedream 5 Lite.
- Text: GPT-6.1 Sol, GPT-6.1 Terra, Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 4 Argon, Gemini 3.5 Flash Lite, Qwen 3.8 Flash, Kimi K3.1 Pro, Kimi K3.1 Pro Highspeed, and Xiaomi's MiMo V2.6 Flash, MiMo V2.6 Pro, and MiMo V2.6 Pro UltraSpeed.
- Embeddings and rerankers: BGE M3, BGE Multilingual Gemma2, BGE Reranker v2 M3, and BGE Reranker v2 Gemma.
- Speech recognition: GigaAM v3, Whisper Large v3, and Qwen3 ASR.
- Text-to-speech: Qwen3 TTS.
Compare prices and specs in the model list.
Export DNS records and add comments to domains
To back up or move a DNS zone, export the domain's records as a .zone file. You can also leave a note on a domain to remember what it's for: click Domain comment under its name. In the domain list, hover over the comment icon to read it.
Jump to any dashboard tab from search
Dashboard search now finds menu sections and their tabs, not just services and docs. Search "ssh", for example, and you'll get Cloud Servers with a link straight to its SSH Keys tab. Matches are highlighted in both section and tab names.
Also Shipped Copy link
App Platform
- Docker and Docker Compose apps can now deploy from any path in the repo, so files like
docker-compose.dev.ymlanddocker-compose.prod.ymlcan live side by side. Set the path in the Project Directory Path field in the Advanced settings section. For a Dockerfile, the path doubles as the build context. - App logs now show domain changes, app stops, and container restarts.
Kubernetes
-
Clusters on older versions now upgrade one minor version at a time, as Kubernetes requires. New clusters use current versions only, but older ones stay available as upgrade steps.
That's it for October. If you need a hand with any of the new features, our support team is here 24/7. See you next month.