Every few weeks, a new "AI agent as a service" launches. Upload your data, connect your tools, and let their agent handle things for you. It's convenient. It's also a trade you should think carefully about before making.
I run self-hosted agents for myself and for clients. I've also tested most of the cloud options. Here's why I keep coming back to self-hosting, and when the cloud option genuinely makes more sense.
The data question
Let's start with the obvious one. When you use a cloud-hosted agent, your data goes to their servers. Your emails, your customer data, your financial records, your internal documents. All of it gets processed on infrastructure you don't control, by code you can't inspect, under terms of service that can change without notice.
"But they say they don't train on your data." Maybe. Today. But you're trusting a promise from a company that might get acquired, might change their privacy policy, or might have a breach. The safest data is data that never leaves your infrastructure in the first place.
With a self-hosted agent running OpenClaw, your data stays on your machines. The agent processes everything locally. The only external calls are to the AI model API (which sees the conversation but not your raw data) and whatever services you explicitly connect.
For businesses handling client data, financial information, medical records, or anything covered by regulations like GDPR or HIPAA, self-hosting isn't just a preference. It's often a requirement.
You control what the agent can do
Cloud agent services decide what tools are available. You get their integrations, their limitations, their permission model. If they don't support a tool you need, you're stuck.
Self-hosted means you control the tool set completely. Need your agent to:
- SSH into a specific server and run a custom script?
- Access a legacy internal API that no cloud service would ever integrate with?
- Read files from a NAS on your local network?
- Control a browser to interact with a web app that has no API?
- Send commands to IoT devices on your network?
With OpenClaw's skill system, you write a skill for it and the agent can do it. No waiting for a vendor to add support. No feature requests sitting in a backlog.
Pricing that makes sense
Cloud agent services charge per interaction, per seat, per automation, or some combination. The pricing always starts reasonable and gets expensive as usage grows. I've seen businesses paying $500+/month for cloud agent services that could run for $100/month self-hosted.
Self-hosted costs are straightforward:
- Hardware or VM: $0 (if you have a spare machine) to $100/month for a cloud VM
- AI model API: $50-200/month depending on usage
- Your time or setup costs: one-time
There's no per-seat pricing. No usage tiers. No "contact sales for enterprise pricing." Run as many agents as your hardware supports. Process as many tasks as your API budget allows.
The Mac Mini setup guide shows how to run a production agent on hardware you buy once and use indefinitely.
Uptime on your terms
When a cloud service goes down, your agent goes down. You have no visibility into the issue, no ability to fix it, and no timeline for resolution. You just wait.
When your self-hosted agent goes down, you can diagnose it immediately. Check the logs, restart the service, fix the config. I've had cloud services take 6+ hours to resolve outages. My self-hosted agents come back in minutes because I can see exactly what's wrong.
You also control the maintenance schedule. No surprise updates that change behavior. No "we're deprecating this feature" emails that force you to redesign your workflows.
Customization depth
Cloud services give you a configuration UI. Checkboxes, dropdowns, maybe a simple scripting language. Self-hosting gives you everything.
Want your agent to behave differently on weekends? Change the config. Want it to use a cheaper model for simple tasks and a smarter one for complex reasoning? Set up model routing. Want custom error handling that sends alerts to a specific Slack channel when a particular skill fails? Write it.
The gap between "what you can configure" and "what you can build" is enormous. Cloud services live in the first category. Self-hosting lives in the second.
The real cost of convenience
The main argument for cloud-hosted agents is convenience. And it's a real argument. Self-hosting requires setup time, some technical knowledge, and ongoing maintenance.
But here's what I've noticed: the businesses that went cloud-first and later switched to self-hosted universally say they wish they'd started with self-hosting. The migration cost (re-building integrations, moving data, retraining workflows) is significant. Starting self-hosted means you never have to make that switch.
The setup cost for self-hosting has also dropped dramatically. With frameworks like OpenClaw, the initial setup takes days, not weeks. And professional setup services bring that down even further.
When cloud-hosted actually makes sense
I want to be fair about this. Cloud-hosted agents are the right choice sometimes:
Quick prototyping. If you want to test whether an AI agent can help your workflow before committing to infrastructure, a cloud service lets you experiment quickly.
Zero technical capacity. If you have no technical team and no budget for professional setup, a cloud service with a good UI is better than nothing.
Temporary needs. If you need an agent for a specific project with a defined end date, spinning up a cloud service is faster than building infrastructure you'll tear down later.
Extremely simple use cases. If you just need a chatbot that answers questions from a knowledge base, cloud services handle this well and self-hosting is overkill.
For everything else, especially ongoing business operations, data-sensitive workflows, or anything that needs deep customization, self-hosting wins.
The security argument
Cloud services are a bigger target than your home server. They hold data from hundreds or thousands of customers. A breach at a cloud agent service exposes everyone's data. A breach at your self-hosted instance exposes only yours.
You also control the security posture. Firewall rules, access controls, audit logging, encryption. All on your terms. For a thorough walkthrough of securing your own deployment, check out the security hardening guide.
What self-hosting looks like in practice
Here's a typical self-hosted setup for a small business:
- A Mac Mini or $40/month VM running the OpenClaw gateway
- Agent connected to Slack for team communication
- Custom skills for email, CRM, and internal tools
- Automated monitoring and restart via launchd or systemd
- Weekly log review and monthly security updates
Total monthly cost: $100-200 including API calls. Total human maintenance: 1-2 hours per month once the initial setup is stable.
Compare that to a cloud service at $200-500/month with less customization, less control, and your data on someone else's servers.
Making the switch
If you're currently using a cloud-hosted agent and considering self-hosting, the process looks like:
- List every integration and workflow your current agent handles
- Set up OpenClaw on your infrastructure
- Rebuild each integration as a skill (or use existing ones)
- Run both in parallel for a week to verify parity
- Cut over and cancel the cloud service
If you want help with this migration, or with setting up self-hosted from scratch, book a call. We've done this transition for several businesses and can make it painless.
The short version: cloud agents are rented convenience. Self-hosted agents are owned infrastructure. For anything you plan to run long-term, ownership wins.
Keep reading: