AI agents read your documents, call your systems and remember what they learn. So the question of where they run is really a question about your data, your costs and your independence. This guide compares SaaS and self-hosted AI agent platforms and ends with a checklist you can take to any vendor.
The short version
| SaaS platform | Self-hosted platform | |
|---|---|---|
| Where it runs | The vendor's cloud | Your servers or your cloud account |
| Who can see your data | You and the vendor | You |
| AI models | Usually the vendor's choice | Any provider, or local models |
| Setup | Minutes | An install by your team or a partner |
| Updates | Automatic | You choose when to upgrade |
| If the vendor changes course | Prices, features or the service itself can change | You keep running what you have |
SaaS wins on speed to start. Self-hosting wins on control. Which matters more depends on what your agents will touch.
When self-hosting matters most
Self-hosting is worth the extra setup when one of these is true:
- Your data can't leave. Regulated industries, government work and some customers' contracts require data to stay in a given country or building.
- You want to choose your models. You may want a cheap model for sorting work, a strong one for analysis, and a local model for anything sensitive.
- Your agents hold credentials. An agent with keys to your accounting or store is safer when the keys stay on your servers.
- You plan to run agents for years. Long-lived memory, skills and history are assets. You want them in a database you control, in a documented format.
Check for hidden lock-in
A platform can run on your servers and still lock you in. Look for these signs:
- a license server the software must contact to keep working;
- phone-home telemetry that is on by default;
- agent logic that lives in the vendor's cloud even when the app is installed locally;
- an undocumented database you can't query with your own tools;
- paid or proprietary components you can't replace.
OrgKeep is built against one test: a customer's developer can clone the repository, run one command and have it in production without contacting the vendor. No license server, no phone-home, telemetry off by default, PostgreSQL with a documented schema, and any AI model through one gateway, including Ollama and vLLM.
Questions to ask any vendor
- Can we run it with no connection to your servers?
- Can we use any AI provider, and local models?
- Do we get the source code, and may we change it?
- Is the database schema documented? Can we export everything?
- Do AI agents have their own identities and permissions, enforced by the server?
- Can an agent use a credential without seeing it?
- Which steps can we make human-only?
- What happens to our installation if we stop paying for updates?
A platform that answers all eight clearly is one you can build on for years.
Where to go next
Read how AI employees work in OrgKeep, how the security model keeps them in bounds, or book a demo to talk through your setup.