Insights
Private and self-hosted AI agents — what actually matters
Local or private deployment does not automatically mean privacy. Practical considerations for self-hosted TypeScript agents.
“Run it on our servers” is a start — not a complete privacy story. Self-hosting the agent runtime does not automatically keep prompts, embeddings, or tool payloads inside your boundary if model providers or SaaS tools sit on the critical path.
What to decide explicitly
- Where prompts and tool payloads are logged
- Which model providers leave your network
- How secrets and tool credentials are stored
- Network isolation between agents and data stores
- Who can approve irreversible actions, and from where
- Retention: how long run history lives, and who can delete it
If those answers are vague, “private AI” is a slogan, not a design.
Topology choices we see most often
| Approach | Fits when | Watch-outs |
|---|---|---|
| Local / on-prem control plane | Strong ownership needs, early builds | Ops burden, model access still TBD |
| VPC-hosted agents + private networking | Cloud estate already mature | Egress to models/tools must be deliberate |
| Hybrid (private data plane, external model) | Speed vs residency tradeoff | Prompt and response data leave the boundary |
There is no single correct topology. There is only a topology that matches your constraints.
HazelJS local-first path
HazelJS emphasizes local development and apply without requiring cloud. That helps ownership of the control plane — it does not by itself guarantee data residency or compliance.
Use local-first to learn the system. Then design production networking and logging with security stakeholders in the room.
What HumanBee will not claim
We do not invent SOC 2, ISO, or residency badges for your deployment. We help you choose deploy topology against real constraints and implement the agent layer honestly inside that topology.
A short checklist before you buy “private”
- Can you name every egress path for model and tool traffic?
- Are logs redacted for secrets and personal data?
- Can you revoke tool credentials without redeploying agents?
- Do approvals work when operators are off the corporate VPN — and should they?