I'll be the first to admit I won't be much help in fixing this stuff. I was troubleshooting with Claude's assistance here and had it write this summary. I'll try to provide follow up information as I can. I ran across this because last night, everything was working fine, and this morning when I fired up the webui, all of my local models (llama.cpp) disappeared.

Summary

get_available_models() in api/config.py performs an SSRF check when fetching models from model.base_url. If the hostname resolves to a private IP and doesn't match the hardcoded allowlist (ollama, localhost, 127.0.0.1, lmstudio, lm-studio), it raises ValueError("SSRF: resolved hostname to private IP") and silently falls back to an empty model list.

This blocks legitimate local-network OpenAI-compatible endpoints like llama-swap, vLLM, llama.cpp, TabbyAPI, etc. from populating the model dropdown.

Reproduction

1. Run an OpenAI-compatible server on the local Docker network (e.g. llama-swap at http://llama-swap:8880/v1)
2. Configure config.yaml:

   model:
     base_url: http://llama-swap:8880/v1
     provider: custom

3. Verify the endpoint is reachable from inside the container:

   docker exec hermes-webui curl -s http://llama-swap:8880/v1/models
   # Returns valid model list

4. Check /api/models — returns {"groups": []}

Root cause

api/config.py ~line 1520:

if not is_known_local:
    raise ValueError(f"SSRF: resolved hostname to private IP {addr[0]}")

The is_known_local check only matches a handful of hardcoded hostnames. Any other container name or local hostname that resolves to a private/link-local IP is rejected.

Impact

• The exception is caught silently (except Exception), so there's no user-visible error — models just don't appear
• The empty result gets persisted to models_cache.json, so the failure is sticky across restarts
• Affects anyone running local inference servers with non-allowlisted hostnames on Docker networks, LAN, etc.

Suggested fix

Either:

• Allow all private IPs when model.provider is custom (the user has explicitly configured a local endpoint)
• Add a config flag like model.allow_private_network: true
• Expand the allowlist to be user-configurable

Environment

• hermes-webui: latest (ghcr.io/nesquena/hermes-webui:latest, pulled 2026-04-26)
• Docker Compose, containers on shared bridge network
• Custom provider: llama-swap (OpenAI-compatible, resolves to 172.18.x.x)