Self-hosted AI
Running AI infrastructure — runtimes, models, or both — on your own servers. Teams self-host for data control, compliance, predictable cost and independence from vendor availability.
What teams are actually buying
Self-hosting is chosen for four reasons that rarely appear in a feature comparison: data never crossing a boundary the compliance team cares about, cost that is capacity-shaped rather than usage-shaped, independence from another company’s incident page, and the ability to keep a model version that a vendor has decided to retire.
The spectrum from runtime to weights
Self-hosting the runtime while calling hosted models is the common first step and solves the credential and audit problem. Self-hosting the weights as well removes the last external dependency and is where regulated environments usually end up. The two decisions are separable and worth making separately.
What it costs you
Hardware, and someone who owns it. GPU capacity sits idle between bursts, models need updating, and the pager is now yours. Self-hosting is the right answer often — but it is an operational commitment, not merely a checkbox that improves privacy.
In AGNT
AGNT ships Docker images for headless self-hosting; pair with Ollama and the entire stack is yours.
Self-hosted AI — common questions
Can AGNT run headless on a server?
Yes. AGNT ships Docker images for headless self-hosting, which is the usual shape for scheduled and webhook-triggered work.
What hardware do local models need?
Small quantised models run acceptably on a modern laptop; larger open-weight models want a dedicated GPU with enough VRAM to hold the weights.
Can I mix self-hosted and hosted models?
Yes, and most teams do — local models for high-volume routine steps, frontier APIs for the calls that need maximum capability.