What this adds

A provider abstraction layer that lets ai-cli work without a Vercel AI Gateway key, by supporting OpenRouter, fal.ai, and direct OpenAI as backends.

Motivation

Getting a Vercel AI Gateway key requires a Vercel account. Many users already have OpenRouter or OpenAI API keys. This change lets them use ai-cli immediately with what they have.

How it works

Auto-detects the active backend from environment variables. Vercel AI Gateway remains the default — existing behaviour is unchanged.

Env vars set	Text	Image	Video
`AI_GATEWAY_API_KEY`	✓ Vercel (100+ models)	✓	✓
`OPENROUTER_API_KEY`	✓ OpenRouter (300+ models)	—	—
`OPENROUTER_API_KEY` + `FAL_KEY`	✓ OpenRouter	✓ fal.ai (Flux, SDXL)	✓ fal.ai (Kling, Wan, Runway)
`FAL_KEY`	—	✓ fal.ai	✓ fal.ai
`OPENAI_API_KEY`	✓ GPT models	✓ DALL-E	—
`AI_PROVIDER=<backend>`	explicit override	—	—

Changes

New file: src/lib/provider.ts

detectBackend() — reads env vars, returns active backend
createProvider() — returns { text(), image(), video() } factory functions pointing to the right AI SDK provider

src/lib/models.ts

fetchGatewayModels(backend) — fetches models from the right source
OpenRouter: live from openrouter.ai/api/v1/models
fal.ai: static catalogue (Flux, Kling, Wan, HunyuanVideo, Runway Gen3) — fal.ai has no public model list endpoint
OpenAI: static catalogue (GPT-4o, o3, o4-mini, DALL-E 2/3, GPT Image 1)
resolveModels(modality, model, known, backend) — selects correct defaults per backend

src/commands/text.ts, image.ts, video.ts, models.ts

All commands call createProvider() and use provider.text/image/video(modelId)
video.ts exits with a clear message if backend has no video support
image.ts language-image path (GPT-4o with image output) is guarded to Vercel gateway only

New dependencies

@openrouter/ai-sdk-provider v2.9.0
@ai-sdk/openai v3.0.58
@ai-sdk/fal v2.0.33

Testing status

TypeScript type-checks clean (tsc --noEmit)
CLI starts and all --help pages render correctly
Provider package APIs verified at runtime (fal.image, fal.video, openrouter() all confirmed callable with correct signatures)
End-to-end generation via OpenRouter (not tested with live key — wiring is correct based on package API verification)
End-to-end generation via fal.ai (not tested with live key — wiring is correct based on package API verification)
Vercel AI Gateway path: unchanged, all existing tests pass

Happy to add end-to-end tests or adjust the approach based on feedback. If the project prefers to keep the tool Vercel-specific, I understand — just thought this was worth proposing given how often it comes up for first-time users.