Skip to content

Models and cost

Zyx does not tie a company to one AI vendor. One runtime runs every model, and a route decides which model does which work.

eve is the one runtime. It runs two kinds of loop:

  • A CLI seat. The Claude or Codex command-line tool runs its own loop, and Zyx’s tools reach it over MCP. A hook asks Zyx’s kernel before every call.
  • eve’s own loop, for every API-key or local model: a local model through Ollama, Bedrock in the EU, Mistral, OpenRouter or the Anthropic API.

Both loops end at the same kernel and the same event log. A route changes where the loop runs, never what governs it.

  • Work picks a level: light (a small model), standard, or deep (the strongest). A failed step retries with more effort first, then one level up.
  • A route can be set per chat, per skill, per skill step, per schedule or per kind of work. The most specific one wins.
  • Flat-rate subscription seats come before paid API calls on the operator’s own install.
  • A client’s install uses no personal subscription. The vendors’ terms keep a personal Claude or ChatGPT login to its owner, so a client’s Zyx uses an API key with a spending limit, an EU-hosted model, or a local model on a server with enough memory.
  • A missing key refuses the turn visibly. A route never moves silently onto a more expensive model.

When privacy matters, a folder or a skill can be routed to an EU-hosted model (Bedrock EU, Mistral) or to a local model (Ollama) that runs on the server. Ollama’s cloud models are refused: local means on this server.

Folder rules cap a folder: which models and providers it may use, dollars per job, dollars per month. The company has a monthly cap too. At the cap, no new work runs until the owner raises it.

  • Langfuse, self-hosted on the server, records every call, its tokens and its scores. Private chats never reach it.
  • The Usage screen shows cost per client folder and per skill.
  • A small judge model grades answers and steps. A down-vote with a note can become a test case.
  • Zyx proposes cheaper models, or a script instead of an AI step, when the numbers show a step always succeeds. The owner decides on the Proposals screen.