Use the club GPU 🔑

Our own AI server, running in-house. Nothing you send it goes to OpenAI, Anthropic, or anyone else. This page takes about two minutes.

Step 1 — get your key

1
Open Buzz and start a direct message with Keymaster 🔑.
2
Send it any message — for example key please.
3
It replies with your personal key, starting with buzzk_. Keep it private.

Your key works only while you're a member here, and only as you. Ask again any time — that replaces your old key with a fresh one.

Step 2 — pick how you want to use it

In Buzz — your own AI assistant (easiest, no terminal)
  1. In Buzz, open Agents in the left sidebar, click New agent, and give it a name.
  2. Under AI configuration choose Customize for this agent, then:
    • Agent harness: Buzz Agent
    • LLM provider: OpenAI-compatible
    • OpenAI API key: paste your buzzk_… key
  3. Open Advanced → Environment variables → Add variable, and add this one (there is no box for it higher up, and without it Buzz talks to OpenAI instead of us):
OPENAI_COMPAT_BASE_URL = https://llm.vibecode.casa/v1
  1. Now set Model to deepseek-v4-flash-dspark. It should appear in the list once the key and the address above are both filled in — if it doesn't, choose Custom model… and type it.
  2. Leave Agent instructions for personality only ("be concise", "you're our research helper") — the reply plumbing is handled for you.
  1. Save, press play to start it, add it to a channel, and mention it by name.

It runs on your own computer while Buzz is open, and thinks on the club GPU.

In opencode

Add this buzz provider to ~/.config/opencode/opencode.json:

"buzz": {
  "npm": "@ai-sdk/openai-compatible",
  "name": "Club GPU",
  "options": {
    "baseURL": "https://llm.vibecode.casa/v1",
    "apiKey": "{env:BUZZ_LLM_KEY}"
  },
  "models": { "deepseek-v4-flash-dspark": {} }
}

Then set export BUZZ_LLM_KEY=buzzk_… and run:

opencode run -m buzz/deepseek-v4-flash-dspark "hello"
In Codex CLI

Add to ~/.codex/config.toml:

[model_providers.buzz]
name = "Club GPU"
base_url = "https://llm.vibecode.casa/v1"
env_key = "BUZZ_LLM_KEY"
wire_api = "responses"

Then:

export BUZZ_LLM_KEY=buzzk_…
codex -c model_provider=buzz -m deepseek-v4-flash-dspark
In Claude Code 
export ANTHROPIC_BASE_URL=https://llm.vibecode.casa
export ANTHROPIC_AUTH_TOKEN=buzzk_…
export ANTHROPIC_MODEL=deepseek-v4-flash-dspark
claude

Works, but Claude Code is built around Claude's own models, so it handles tools less reliably here than opencode or Codex do.

Which model?

Use deepseek-v4-flash-dspark unless you know you want something else. nemotron-super and laguna-s-2.1 are also available when running.

If something doesn't work