🔒 Air-gapped, self-hosted agents
Stack: Deploy: Docker Compose / Kubernetes (your infra) · Models: local (Ollama, vLLM) · Secrets: your vault · State: your Postgres · Output: agents, nothing leaves the network
Who it’s for: teams handling records that legally cannot leave the network, patient data, case files, trading positions. What it does: runs the whole platform on your own infrastructure against a local model, so agents work end to end without a single call crossing the boundary. Because the core is open source (Apache-2.0), this is configuration, not a special edition, the hosted service runs the same code.
Build it
-
Self-host the platform with Docker Compose (single box) or the Helm chart (Kubernetes). Postgres and object store are yours; no telemetry leaves the cluster.
-
Point at a local model: add an AI Model whose base URL is a server you run (Ollama, vLLM, any OpenAI-compatible endpoint) inside the network. Inference stays on your hardware.
-
Keep secrets and data in place: keys in your own Credentials vault, memory in your Postgres, tools executed by a runner on machines you control.
Make it yours
- The same governance, verify gates, approvals, audit, applies unchanged, just inside your walls.
- Use a runner to reach on-prem systems the cluster can’t (Give an agent a machine).
- Nothing has to cross the boundary: model calls, tool calls, secrets, and state all stay internal.
Uses: Enterprise · Self-host · Compliance