--- title: "Where to Host an Autonomous AI Agent 24/7 (Options Compared)" description: "VPS, cloud VMs, managed platforms — compare every option for hosting an autonomous AI agent 24/7, with state persistence, scheduling, and included AI credits." canonical: https://agentroost.app/en/blog/where-to-host-autonomous-ai-agent-24-7 date: 2026-05-17T04:00:00Z --- [Canonical URL](https://agentroost.app/en/blog/where-to-host-autonomous-ai-agent-24-7) # Where to Host an Autonomous AI Agent 24/7 (Options Compared) Most tutorials end with `python agent.py` running in a terminal. That works for ten minutes — until your laptop sleeps, your SSH session drops, or the process exits on an unhandled exception and nobody notices until Monday morning. Running an autonomous AI agent 24/7 is a different problem. Before you pick a platform, you need to understand what "always-on" actually requires. --- ## The Four Things a 24/7 Agent Actually Needs ### 1. Uptime and crash recovery A process supervisor (systemd, supervisord, Docker `restart: always`) must restart the agent if it dies. Without one, any uncaught exception or OOM event kills your agent permanently until someone SSHes in. ### 2. State that survives restarts An agent with no memory is a stateless chatbot. A *persistent* agent remembers what it did yesterday — open tasks, conversation history, the last price it fetched, the draft it was writing. This means either a database, a volume-mounted file store, or a framework that handles persistence for you. ### 3. A scheduler Autonomous agents do things *unprompted* — check a feed every hour, send a daily digest, run a report every Friday. You need cron, a task queue, or a built-in scheduling primitive. A sleeping process with a `time.sleep(3600)` loop is not a scheduler; one SIGKILL and the timing drifts. ### 4. An AI bill you can predict Every LLM call costs money. If you wire your agent to OpenAI or Anthropic directly, the bill grows with usage and arrives as a surprise. Many people discover this after their overnight research agent racks up $40 in a single run. --- ## Option 1: Self-managed VPS (Hetzner, DigitalOcean, Vultr, etc.) You rent a server, write the agent code, and manage everything yourself. **What you get:** Full control. Any language, any framework, any model via the provider's API. **What you manage:** - SSH into the box, clone your repo, write a systemd unit or `docker-compose.yml`. - Handle log rotation, out-of-memory restarts, and kernel updates. - Mount a persistent volume or set up a database for state. - Add cron entries or a task queue for scheduling. - Pay your LLM provider separately — OpenAI, Anthropic, OpenRouter, etc. — and set spending limits manually. - Renew SSL certs, keep dependencies up to date, rotate API keys when they leak. **Reality check:** A $6/mo Hetzner CAX11 is cheap, but you still need to add your own API key (easily $10–30/mo for moderate agent usage), configure a process supervisor correctly, and debug the 2 a.m. restart loop. Total cost is real but unpredictable. Time cost is also real — initial setup is 2–4 hours, ongoing maintenance is 30–60 minutes a month *when things go right*. This is the right choice if you are already comfortable with Linux system administration and you need capabilities (custom GPU, exotic networking, a specific OS image) that a managed platform does not offer. --- ## Option 2: Cloud Functions / Serverless (AWS Lambda, Google Cloud Run, etc.) Serverless works brilliantly for stateless, event-driven tasks. It does not work well for autonomous agents. **The fundamental mismatch:** Cloud Functions time out after seconds to minutes. A long-running research agent that takes 10 minutes to finish a task will be killed mid-run. State does not persist between invocations unless you bolt on a database. Scheduling via EventBridge or Cloud Scheduler adds another layer to configure. You still pay your LLM provider separately. Serverless is excellent for *triggering* an action (a webhook fires, an email arrives). It is a poor host for an agent that thinks for extended periods or maintains ongoing context. --- ## Option 3: Kubernetes / Container Orchestration If you already run Kubernetes, you can deploy an agent as a long-running Deployment with a readiness probe and a restart policy. Persistent state goes in a PVC. Scheduling happens via a CronJob resource. This scales well, but the operational overhead is substantial. Kubernetes is a platform for teams that operate many services. Running a single AI agent on a personal K8s cluster is engineering overkill for most use cases. --- ## Option 4: Managed Agent Platforms A handful of newer platforms handle the infrastructure layer for you. The key question to ask any of them: **does AI usage come included, or do I bring my own API key?** | Platform | What it hosts | LLM credits included? | Process management | Persistent state | Built-in scheduler | |---|---|---|---|---|---| | **AgentRoost — Hermes** | Always-on persistent AI agent | **Yes, included** | Supervised, auto-restarts | Yes, cross-day memory | Yes, built-in | | Elestio / Sliplane | Generic Docker apps | No (BYOK) | Docker restart policy | Via mounted volumes | Manual cron | | n8n Cloud | n8n SaaS (shared) | No (BYOK) | Managed by n8n | Workflow state only | Trigger-based | | Railway / Render | Long-running containers | No (BYOK) | Restart on crash | Via volumes or DB add-on | Via cron add-on | "BYOK" (bring your own key) means you wire your OpenAI or Anthropic key yourself and the LLM bill arrives separately. On AgentRoost, AI credits are part of the subscription — the agent works out of the box, no key to paste, no separate bill to manage. --- ## How Hermes Solves the Four Requirements **Hermes** is the agent framework on AgentRoost designed specifically for always-on, autonomous operation. Here is how it maps to the four requirements above: - **Uptime:** The runner process is supervised and auto-restarts on crash. You do not have a systemd unit to write. - **State persistence:** Conversation history, task state, and memory are stored across restarts by default. Your agent remembers what it was doing when it wakes back up. - **Scheduling:** Hermes supports scheduled tasks natively. You tell the agent to run a task every morning and it runs every morning — no cron tab, no external task queue. - **LLM bill:** AI credits are included in the subscription. You access 350+ models and switch between them without touching an API key. The cost is predictable because it is folded into a flat monthly price. The interface is a Telegram bot provisioned automatically when you create the agent. There is nothing to configure — you open the manager bot, hit `/start` on your agent, and you are talking to it within two minutes. --- ## Running Your Own Hermes Agent on AgentRoost 1. **Sign up** at agentroost.app with email, Google, Microsoft, or Discord. 2. **Pick the Hermes framework** from the agent catalog ([/en/agents/hermes](/en/agents/hermes)). 3. **Name your agent.** The runner spins up on dedicated hardware. 4. **Open the AgentRoost manager bot on Telegram**, then `/start` your new agent — it introduces itself. 5. Give it a scheduled task: *"Every weekday at 8 a.m., summarize the top three posts from [RSS feed URL] and send them here."* The agent runs it tonight, and every weekday after that, whether your laptop is on or not. Pricing starts at **$19.99/mo all-in** — compute, uptime management, AI credits, Telegram provisioning. [Compare plans](/en/pricing) to see if it fits your budget. There is a 14-day money-back guarantee, billed monthly, cancel anytime. --- ## Which Option Should You Pick? | Your situation | Best fit | |---|---| | You want zero DevOps, scheduled tasks, cross-day memory, no API key wrangling | **Hermes on AgentRoost** | | You need root access, custom hardware, or a workflow-automation GUI | Self-managed VPS or [your own n8n instance](/en/agents/n8n) | | You already run Kubernetes and want consistency with your stack | K8s Deployment + CronJob | | Your agent is purely event-driven and stateless | Cloud Functions (Lambda, Cloud Run) | The honest answer: for most people who want an agent that "just runs," the DevOps path costs more in time than it saves in money. Managed platforms eliminate the setup cost; AgentRoost also eliminates the LLM billing complexity. --- ## Tips Before You Commit to Any Platform - **Test with a short scheduled task first.** If a platform cannot reliably fire a cron job, it will not reliably host your agent. - **Check what "state persistence" means specifically.** Some platforms reset the container on restart; your in-memory state is gone. Ask whether files on disk or a database survive a process crash. - **Model-switching matters for autonomous agents.** A research agent might want a large context window; a daily digest bot is fine with a smaller, cheaper model. Make sure you can switch without redeployment. - **Predict your LLM cost before going BYOK.** If your agent makes many LLM calls per day, that cost compounds fast and lands as a surprise invoice. With included credits, the math is simpler: you know the ceiling when you sign up.