Skip to content

Self-hosting

AI models and costs

Use an OpenRouter key or your own OpenAI-compatible model, and see the cost before ZeroWidth reads anything.

View as Markdown

ZeroWidth uses an AI model to read decisions, answer questions and dream. Each of those calls costs money with a hosted model.

OpenRouter

The default. ZeroWidth reads your key from OPENROUTER_API_KEY, or from ~/.config/zerowidth/env. To use another environment variable, name the variable in zerowidth.json as model.keyEnv. You pay OpenRouter directly. ZeroWidth only uses model hosts that keep nothing and don't train on what's sent.

Your own model

ZeroWidth works with any OpenAI-compatible endpoint, such as Ollama, vLLM or LM Studio. With your own model on your own network, nothing leaves your network.

Terminal
zerowidth serve --db team.db --endpoint http://localhost:11434/v1 --model llama3.1 --embed-model nomic-embed-text

A key for the endpoint, if the endpoint needs one, goes in ZEROWIDTH_API_KEY.

Spending limits

import, reversals and metrics --propose are dry runs first: each shows the cost by model and reads nothing until you add --run --cap <dollars>, which reads up to that limit and stops.

Other commands spend as they run, with no estimate and no limit: ingest, settle, dream, duplicates and ask, and Dream now in the console. Use your model provider's own spending limit as well.