Reference
Models and endpoints
Base URL
https://llm-34a13a96.bunsenbrenner.org
Model
| LiteLLM model name | Backing model | Where it runs |
|---|---|---|
local-qwen3-coder |
qwen3-coder:30b (MoE, ~3.3B active params, Q4_K_M, ~18GB on disk) |
Ollama, this origin’s NVIDIA TITAN RTX (24GB VRAM) |
Other model names configured on this same LiteLLM instance (cf-llama-70b, cf-llama-free, and
the *:cloud fallback chain) are cloud-routed and not accessible to keys scoped to
local-qwen3-coder — see Add an authorized person.
Endpoints
| Path | Shape | Notes |
|---|---|---|
POST /v1/chat/completions |
OpenAI | Standard chat completions. |
POST /v1/messages |
Anthropic | LiteLLM’s Anthropic-format passthrough; what Claude Code talks to. |
GET /health/liveliness |
— | No auth required; used by the tunnel’s own health checks. |
measured Verified live 2026-08-29: /health/liveliness answers 200 ("I'm alive!") unauthenticated; /v1/chat/completions and /v1/messages both exist and return a 401 auth error when called with no key.
Auth
Authorization: Bearer sk-... (OpenAI-shaped calls) or x-api-key: sk-... +
anthropic-version: 2023-06-01 (Anthropic-shaped calls). Same key works for both.
Known deployment-specific quirks
drop_params: trueis set proxy-wide, because Claude Code sends acontext_managementparameter Ollama’s backend doesn’t understand; without it every Claude-Code-originated request fails with a 400.- not built The model’s context window is advertised as 256K by its card, but this deployment has not raised
Ollama’s default
num_ctx— treat anything past roughly a few thousand tokens of real context with suspicion. Seecapability-probe/transcripts/05-long-context-needle.jsonin the code repo for the exact failure.
Found an error, or something that didn't work as documented? Open an issue →