Documentation

Nebius Token Factory serves open models over an OpenAI-compatible API. It does not speak the Anthropic Messages API that Claude Code uses, nor the OpenAI Responses API that Codex uses. The relay runs a small local daemon that translates those wire formats to Nebius /chat/completions on the fly - your agent believes it is talking to its native backend, while every token is served by Nebius.

The one-liner installs nebiusrelay plus a short alias per harness into ~/.nebiusrelay/bin/, and installs Bun for you if it is not already present.

First run walks you through configuration, or run it directly:

You will be asked for two keys:

Launch any of them directly. Extra arguments are passed straight through to the underlying agent.

The released CLI includes nebiusrelay chatgpt (alias: codex-app) for compatible desktop coding tasks. It configures a local Responses provider in ~/.codex/config.toml, writes a model catalog, and attempts to open the desktop app.

Restore the backed-up configuration when you are finished:

Models powered byNebius Token Factory

The model list is fetched live from Nebius at startup, so every model they serve is available and each model's vision support comes from the API's own modality field - never a hand-maintained list. Results are cached locally and fall back to a bundled snapshot when offline.

Relay 0.15.4 adds bundled support for GLM 5.3 and DeepSeek V4 Pro 0813. GLM 5.3 Flash remains the default. Select either new model explicitly:

nebiusrelay --model zai-org/GLM-5.3 codex
nebiusrelay --model deepseek-ai/DeepSeek-V4-Pro-0813 codex

In Relay 0.15.3 and later, select a vision-capable model explicitly when launching Codex. The default GLM model is text-only. Image options belong to Codex, not every harness; other agents keep their native attachment controls.

Use Codex's clipboard-image shortcut where your terminal supports it. Normal terminal paste may paste text only; the relay does not monitor your clipboard. For a reliable file-based alternative, save the image on the machine running Codex:

Web search powered byTavily

Claude Code and Codex expose a native web_search tool. The relay backs it with Tavily: with a key configured, searches return real results with citations. Without one, a search returns a clear "TAVILY_API_KEY not set" message rather than failing silently. Nebius has no hosted search tool, so this is how agents get live web access.

Claude and Codex are proxied, so the daemon meters every turn. The other harnesses hold the key and call Nebius directly, which is why they report $0.00. NEBIUSRELAY_METER=1 points them at the daemon instead:

Nebius TF Relay ▸ Launching Pi Code with Nebius Token Factory.
[nebiusrelay cost] session total: $0.0056 (1,518 in, 69 out)

They then share the same client as everyone else - automatic model fallback, the per-model circuit breaker, transient-fault retries - and the real Nebius key stays inside the daemon, since the harness only ever sees a local session token. If the daemon is unreachable the launcher says so and connects directly, so metering can never be the reason a session fails to start.

An LLM-readable doc is published at llms.txt. If you are an agent asked to install, configure or drive nebiusrelay - including headless - read that first. It covers install, configure, every command, the models, and headless usage patterns.