AI providers

FX runs a workflow's own code with no provider at all. A step that calls an AI model (an image, a clip, a 3D model, structured text) runs on a route, model@provider, and a live run needs that provider's key, or its program for Codex. This page lists the ways to set one up and what each one serves.

Path	What you set up	What it serves	Who bills
No provider	nothing	your own code and the built-in local steps	nobody
Your own Codex	Codex, signed in to ChatGPT	images	your ChatGPT plan; $0 to FX
Grida Gateway	GG_API_KEY	text, agent turns, images, video, 3D models and rigs	your organization's Grida AI credits
Your own keys	each provider's key	that provider's routes	each provider's account

You can mix them: a project picks a route per capability in fx.yaml, and a step can name its own route:. grida fx doctor says which keys FX found, where, and which routes a live run could serve (Running).

No provider

FX is a workflow engine first. Your Python nodes, the built-in local steps (image.resize, json.merge, package, …) and file facts run with no key and without --live, and a run of them is cached and recorded like any other. Without --live, a step that would call a provider is refused for nothing, so you can build and test the rest of a workflow first, or answer paid calls with stand-ins.

Your own Codex

With Codex installed and signed in to your ChatGPT account, the route imagegen@codex draws and edits images with no API key. Codex counts each image against your plan; FX books it at $0, and a run whose calls are all on Codex needs no --max-usd.

It serves images only (image.generate and image.edit), and Codex picks the model and the quality. Setup and caveats, sizes above all: Images from your own Codex.

Grida Gateway

One key serves every route on Grida Gateway (@gg). Create it at console.grida.co, where you also add credits and see what each call was charged, then set it:

GG_API_KEY=…

in the environment, or in a .env file in your project (keep it out of version control). FX books a GG call at what GG charged for it (providers.md §9.7).

The key needs both scopes, text for text and agent steps and media for everything else. GG's own API reference, written for people and agents alike, is gateway.grida.ai/llms.txt: key scopes, credits, spending caps and limits.

GG serves text and agent turns, images, video, and 3D models and rigs. It does not serve sound, speech or music: those routes need their provider's own key (below).

Your own provider keys

Every other route takes its provider's key under its usual name. Set only the ones your routes use; a missing key refuses only the calls that need it, before anything is sent.

Provider	Variable	Built-in routes for	Get a key
OpenAI	OPENAI_API_KEY	images	platform.openai.com
OpenRouter	OPENROUTER_API_KEY	text, agent turns, images, music	openrouter.ai
fal	FAL_KEY	images, video	fal.ai
Tripo	TRIPO_API_KEY	3D models and rigs	platform.tripo3d.ai
ElevenLabs	ELEVENLABS_API_KEY	sound effects, speech	elevenlabs.io

OPENAI_BASE_URL, OPENROUTER_BASE_URL, FAL_BASE_URL, ELEVENLABS_BASE_URL and GG_BASE_URL point a provider at another address, such as a proxy. How FX reads keys and the .env file: Keys and live runs.

The built-in routes

Capability	Grida Gateway	Codex	Your own keys
structured.generate, agent.turn	✓		OpenRouter
image.generate, image.edit	✓	✓	OpenAI, OpenRouter, fal
video.generate	✓		fal
mesh.generate, mesh.rig	✓		Tripo
music.generate			OpenRouter
sound.generate, speech.generate			ElevenLabs

The routes themselves, with their models and planning prices, are in FX's built-in route table: grida fx nodes <type> lists a capability's routes, and grida fx doctor says which ones a live run could serve. A project can add or restate routes in its own route_tables: (where prices come from).