LiteLLM¶
Primary audience: TAPPaaS admin and users of AI models (via apps or the API).
Unified AI API gateway — routes requests to multiple LLM providers with usage tracking, caching, and access control.
What you get¶
| Capability | Access from | How |
|---|---|---|
| Unified LLM API | Any internal zone (via pinhole) | OpenAI-compatible endpoint on port 4000 |
| Web UI | Internal network | http://litellm.srvWork.internal:4000/ui |
| Usage tracking | Admin UI | Per-key request counts, cost, latency |
| Response caching | Automatic | Redis-backed; reduces provider API costs |
| Virtual API keys | Admin UI | Scope per user or application |
| Model service for modules | litellm:models consumers | Virtual key auto-provisioned at install |
What is not included¶
- Provider API keys (added via UI post-install — see INSTALL.md)
- Model selection (configured in UI after install)
- External access — internal zones only; expose via dmz if needed
Requirements¶
srvWorkzone (VLAN 220)- NixOS template (
templates:nixos) - 4 vCPU, 4 GB RAM, 32 GB disk by default (sizing guidance in DESIGN.md)
Operations¶
Provider credentials (OpenRouter, Perplexity, Abacus, etc.) are managed as named, referenceable objects — models point at a credential by name rather than embedding the key, so rotating a key updates every model that uses it in one step.
| Task | Command |
|---|---|
| See models, credentials, virtual keys, teams (no secrets shown) | scripts/litellm-credentials.sh inspect |
| Register a new provider credential | scripts/litellm-credentials.sh add --name <name> --provider <provider> |
| Rotate an existing credential's value | scripts/litellm-credentials.sh rotate --name <name> |
| Wire a model to a credential | scripts/litellm-credentials.sh assign-model --model <model> --credential <name> |
| Full end-to-end key rotation (env file + credential + restart) | scripts/rotate-provider-key.sh --vmname <vmname> |
Dependencies¶
| Depends on | Purpose |
|---|---|
cluster:vm | VM provisioning |
templates:nixos | NixOS base image |
backup:vm | Daily backups |
identity:identity | Secrets management |
network:proxy | HTTPS reverse proxy |
network:rules | Internal firewall pinholes |
vllm-amd:inference | Local LLM inference backend |
For installation steps see INSTALL.md.