fx-openai

A fork of vercel-labs/fx with OpenAI-compatible API support.

Use any OpenAI-compatible endpoint (OpenAI, OpenRouter, Ollama, LiteLLM, local proxies, etc.) instead of being locked to Vercel AI Gateway.

Quick start

curl -fsSL https://fx.sh/setup.sh | bash   # or build from source below

export OPENAI_API_KEY="your-api-key"
export FX_OPENAI_BASE_URL="https://api.openai.com/v1"   # optional; this is the default

fx models          # fetches models from {base}/models
fx ask "hello"     # uses {base}/chat/completions

Custom endpoints

# OpenRouter
export OPENAI_API_KEY="sk-or-..."
export FX_OPENAI_BASE_URL="https://openrouter.ai/api/v1"

# Ollama (OpenAI-compatible)
export OPENAI_API_KEY="ollama"
export FX_OPENAI_BASE_URL="http://127.0.0.1:11434/v1"

# Any compatible proxy
export OPENAI_API_KEY="your-key"
export FX_OPENAI_BASE_URL="https://your-proxy.example/v1"

You can also use FX_OPENAI_API_KEY instead of OPENAI_API_KEY.

Set a model explicitly when your provider uses plain model IDs (not provider/model Gateway IDs):

export FX_MODEL="gpt-4o"
fx ask "Summarize this repo"

How it works

When OPENAI_API_KEY (or FX_OPENAI_API_KEY) is set, fx-openai:

1. Authenticates with your OpenAI-compatible API key (takes precedence over Vercel Gateway login)
2. Lists models via GET {FX_OPENAI_BASE_URL}/models
3. Runs the agent via POST {FX_OPENAI_BASE_URL}/chat/completions with streaming SSE

Vercel AI Gateway mode still works when no OpenAI key is configured (fx login, AI_GATEWAY_API_KEY, etc.).

Build from source

Requires Zig 0.16.0+:

git clone https://github.com/iamdemetris/fx-openai.git
cd fx-openai
zig build -Doptimize=ReleaseSafe
./zig-out/bin/fx status

Environment variables

Variable	Purpose
OPENAI_API_KEY	API key for OpenAI-compatible providers
FX_OPENAI_API_KEY	Alias for OPENAI_API_KEY
FX_OPENAI_BASE_URL	Base URL (default: https://api.openai.com/v1)
FX_MODEL	Override the model for one run

All other fx settings and commands work as upstream.

Upstream

Based on vercel-labs/fx (Apache-2.0).

Changes in this fork:

• OpenAI-compatible provider mode (src/core/config/openai_config.zig, src/core/gateway/openai_json.zig, src/gateway/openai_client.zig)
• Custom remote endpoint support when OpenAI mode is active
• Model catalog fetch from /v1/models

License

Apache-2.0 — see LICENSE.