← Home

How it works

Clone it, point it at a model you own, run it. Every drop reads the same three variables, so what you set up for the first one works for the hundredth.

  1. 01

    Clone the one you want

    Every drop is its own repository. Its own compose file, its own dependencies, its own README — nothing shared, nothing to check out that you didn't ask for.

  2. 02

    Point it at a model

    The same three variables in every drop. A free hosted key takes about a minute and needs no card, or run fully local against Ollama and nothing leaves your machine.

  3. 03

    docker compose up

    That is the install. If it takes longer than a minute, that is a bug worth reporting — not a step you were supposed to figure out.

One config

.env
LLM_BASE_URL=https://openrouter.ai/api/v1
LLM_MODEL=google/gemma-4-26b-a4b-it:free
LLM_API_KEY=                    # free key, no card

Anything that speaks the OpenAI chat-completions shape works. Point LLM_BASE_URL at Anthropic and drops that support it switch wire format themselves rather than making you configure it.

Or fully local

shell
brew install --cask ollama
OLLAMA_CONTEXT_LENGTH=16384 ollama serve
ollama pull qwen2.5-coder:7b

OLLAMA_CONTEXT_LENGTH is not optional. Ollama defaults to 4096 tokens and truncates longer prompts silently — generation looks fine while editing appears completely broken.

Free providers, no card

ProviderBase URLKey
OpenRouterhttps://openrouter.ai/api/v1Get one ↗
Groqhttps://api.groq.com/openai/v1Get one ↗
Cerebrashttps://api.cerebras.ai/v1Get one ↗
Google AI Studiohttps://generativelanguage.googleapis.com/v1beta/openaiGet one ↗
NVIDIA NIMhttps://integrate.api.nvidia.com/v1Get one ↗

Read this first

  • There is no keyless free model. Every hosted provider rejects unauthenticated requests — checked, not assumed. Free means free of charge, not free of signup.
  • Free tiers are smaller than they look. Measured: OpenRouter gives 50 requests a day on :free models, Google AI Studio gives 20 a day on gemini-2.5-flash. A 10-slide deck spends eleven.
  • A daily cap is not a rate limit. Google answers a daily quota with a 26-second retry delay, which is useless advice — the limit resets at midnight. Drops tell the difference and fail fast instead of backing off for three minutes.
  • You do not need a GPU. The default is a free hosted model. Nothing here assumes a frontier model.

Something you want replaced?

The best nominations are specific: what it costs, what it actually does underneath, and the exact moment the paywall ruined your day.

Nominate a target