AI agent hosting.

Option 1: a local machine

A laptop, a desktop, or a Mac Mini in a cupboard. Free compute you already own, and the honest starting point for almost everyone. It stops being viable at the exact moment the agent starts mattering.

• Good for: development, experimentation, and agents that only need to run while you are working anyway.
• Breaks on: sleep, reboots, power cuts, travel, and network changes.
• A Mac Mini that stays on solves the lid problem and none of the others: you still patch it, restart the agent, rotate its keys, and have no alerting.

Option 2: your own VPS

The most common next step, and a genuinely reasonable one if you like servers. Roughly $12 to $18 a month buys a machine that stays on. What it does not buy is anything above the operating system.

• Good for: full control, unusual dependencies, strict data residency, and people who already run servers.
• You operate: hardening, the runtime install, updates, restart policies, log rotation, backups, and alerting.
• Honest cost: the invoice is $12 to $18. Setup runs to a few hours and upkeep to a few hours a month, which dominates the total for most people.

Option 3: serverless

Functions that run on a schedule and bill per execution. Excellent for a narrow shape of agent and awkward for the rest, and the mismatch is structural rather than a matter of configuration.

• Good for: short, stateless, scheduled jobs with predictable runtimes.
• Fights you on: long-running work against execution timeouts, persistent state between runs, warm interactive sessions, and anything wanting a filesystem.
• Cost is genuinely low when the fit is right, and the fit is right less often than it first appears.

Option 4: a managed agent host

A platform whose job is running agents rather than running machines. Higher monthly price, and the comparison is only meaningful if you count what it replaces.

• You operate: the agent's instructions, tools, and schedule.
• The platform operates: provisioning, isolation, restarts, health monitoring, secret storage, run records, and spend limits.
• Good for: anyone whose interest is the agent's output rather than its infrastructure.

How to choose, in one question

Ask what you want to be doing on a Sunday when an agent stops. If the honest answer is diagnosing it, self-host and enjoy it. If the honest answer is not finding out until Monday, you need something that restarts the agent and tells you it did.

Then, and only then, pick a runtime

Hosting and runtime are separate decisions taken in that order. Once you know where the agent should live, choose what it runs on: OpenClaw for the OpenClaw operating model and existing setups, Hermes for config-driven agents with schedules, gateways, and approvals, Codex for code-first agents on the OpenAI Codex CLI. Qoren runs all three in the same managed environments.