Introduction
Hyphen Developers
Hyphen is one flat-rate key for every frontier model, behind one base URL.
Point any OpenAI- or Anthropic-compatible tool at the Hyphen gateway, drop in your key, and you are done. No juggling provider accounts, no per-provider billing, no surprise overage — just three stable model aliases and a monthly cap you control.
Code
One base URL, one key
Every endpoint — Chat Completions, Responses, Anthropic Messages, embeddings — is served from the same base URL and authenticated with the same key.
Why it works everywhere
Under the hood the gateway is an OpenAI-compatible proxy (built on LiteLLM). It speaks three wire formats so your tools don't have to care what runs upstream:
| Endpoint | Wire format | Used by |
|---|---|---|
POST /v1/chat/completions | OpenAI Chat Completions | The default surface — OpenAI SDK, Cursor, Cline, Aider, Continue, Hermes |
POST /v1/responses | OpenAI Responses API | Codex CLI |
POST /v1/messages | Anthropic Messages API | Claude Code |
GET /v1/models | Model discovery | Any client |
POST /v1/embeddings | OpenAI Embeddings | RAG / search |
Because the gateway translates between these formats, a tool that only speaks the Responses API (Codex) or only speaks Anthropic Messages (Claude Code) can still reach the same upstream models as everything else.
The three model aliases
| Alias | Reach for it when… |
|---|---|
hyphen-sonnet | Balanced default for day-to-day agent work |
hyphen-opus | Hardest reasoning, large refactors, deep planning |
hyphen-fast | Cheap, low-latency, high-volume calls |
Aliases are stable. The upstream model each one maps to is managed inside the gateway, so your configs keep working as models are upgraded.
One key, one cap
Each key carries a monthly USD budget. When it is spent, requests return HTTP 429 with the reset date — nothing runs over budget by accident. See Rate limits & caps.
Next steps
- Quickstart — get a key and make your first request.
- Connect your coding agent — Claude Code, Codex CLI, Hermes, OpenClaw, and more.
- API Reference — the full interactive spec.