airmetDocsJoin the pilot
Docs

Two legs, three ways to get them, one binary on every channel.

Airmet is one sidecar: change one base URL and it races a second copy of a share α of your LLM calls (about 30% by default), streaming whichever token arrives first. By default the other leg is not cancelled for being slower: it runs to its own finish unless the runaway guard stops it for looping or running far past the winner, so its prompt cache stays warm. Every doc below hangs off the same three setups — start with the two providers you can reach, and move between setups with a config change, not a rewrite.

Setup 1 · one marketplace key, two pinned providersA pinned pair on one accountEach leg pinned to a different underlying provider on a major model marketplace: independence without a second account. The shape our pilot runs.
Setup 2 · two provider keysTwo providers, one leg eachKeys for two providers that serve the same model — the same config file, one key per leg.
Setup 3 · one key, the same provider twiceA provider raced against itselfThe fallback when you hold one key: two draws from one provider's queue. The shape of run R1 — it cuts the provider's own variance, not a whole-provider bad hour.

Which setup am I?

All docs

Install channels: pip, npm, Docker and a bare download carry the same binary, one build per platform; vLLM / SGLang, LiteLLM-as-upstream, n8n, MCP and the IDE quickstart are covered in the install-channels guide.

Running a pilot with us? Join the pilot and we set it up with you.

No analytics until you say so. One first-party cookie remembers this choice; allowing adds EU-hosted usage analytics and error reporting (what we send); declining changes nothing about how the site works.