Self-hosted AI agents: choose the boundary before choosing the stack
Compare self-hosted AI agent deployment boundaries for models, runners, data, connectors, control planes, observability, and enterprise operations.
Self-hosted AI agents are not a single yes-or-no feature. Decide separately where models, workflow runners, data stores, credentials, connectors, traces, and the control plane run. The correct boundary follows data residency, network, latency, assurance, and operational requirements.
Break deployment into components
A local model does not make the whole agent local. Inventory model inference, embeddings, retrieval stores, workflow execution, secrets, business-system connectors, trace storage, updates, and administration. Draw every network boundary and identify which component can see sensitive content.
Choose an operating model
Common patterns include local models with a hosted control plane, customer-hosted runners with managed orchestration, and a fully private deployment. Each changes who patches infrastructure, handles backups, monitors abuse, rotates credentials, and responds to incidents.
- Data residency and egress requirements
- Inbound and outbound network policy
- GPU and capacity ownership
- Upgrade and vulnerability response
- Backup, disaster recovery, and evidence retention
Melaya's current boundary
Melaya supports local model providers and local execution paths. Fully private or customer-specific deployment is a Citadel architecture discussion. Do not assume every control-plane component is customer-hosted without a written deployment design and responsibility matrix.
Frequently asked questions
Is a local LLM the same as a self-hosted agent platform?
No. The model may run locally while orchestration, retrieval, credentials, connectors, or traces remain hosted elsewhere.
When is self-hosting worth it?
It is most justified by explicit residency, network isolation, customization, latency, assurance, or cost requirements that outweigh the added operational burden.
Can Melaya use Ollama or LM Studio?
Yes. Melaya supports local-model workflows, subject to the capabilities and deployment configuration of the selected plan.
