AI Gateway

ai-gateway CLI

A thin wrapper around the ai-gateway CLI (https://vercel.com/ai-gateway) for one-shot text, image, and video generation. Use it whenever a single CLI call beats writing SDK code.

Defaults

• Text model: xai/grok-4.1-fast-non-reasoning (cheap + capable)
• Image model: google/imagen-4.0-fast-generate-001
• Video model: xai/grok-imagine-video
• Override per-call with -m <model-id>. Browse with ai-gateway models --type image|language|video.

Notable models

• Image-only: bfl/flux-2-pro, bfl/flux-pro-1.1, openai/gpt-image-2, google/imagen-4.0-generate-001, xai/grok-imagine-image.
• Multimodal LLMs (auto-routed through chat completions): google/gemini-2.5-flash-image (Nano Banana), google/gemini-3.1-flash-image-preview (Nano Banana 2), google/gemini-3-pro-image.
• Text quality: anthropic/claude-opus-4.6, openai/gpt-5.4, xai/grok-4.3.
• Video: xai/grok-imagine-video (cheap), bytedance/seedance-v1.0-lite-t2v, google/veo-3.1-fast-generate-001 (premium, audio), klingai/kling-v2.6-t2v.

Steps

1. Verify the CLI is installed

Run which ai-gateway. If missing, install: npm install -g @vesely/ai-gateway-cli.

Success criteria: which ai-gateway returns a path.

2. Verify the API key is reachable

The CLI looks for the key in this order: --key flag → AI_GATEWAY_API_KEY env → ~/.config/ai-gateway-cli/config.json. If AI_GATEWAY_API_KEY is unset AND the config file is missing/empty, ask the user for it once: tell them to either export AI_GATEWAY_API_KEY=... or run ai-gateway config set key <value>. Get a key at https://vercel.com/ai-gateway.

Success criteria: ai-gateway config shows a key (masked) OR $AI_GATEWAY_API_KEY is set.

3. Run the generation

Text (streamed to stdout):

ai-gateway "<prompt>"                       # default model
ai-gateway -m anthropic/claude-opus-4.6 "<prompt>"
ai-gateway --json "<prompt>"                # full JSON response (.text, .usage)
cat file.md | ai-gateway "<prompt>"         # piped stdin is prepended as context

Image (saves to disk, prints path):

ai-gateway image "<prompt>"                                    # ./ai-image-<timestamp>.png
ai-gateway image -o output.png "<prompt>"                      # custom path
ai-gateway image -n 4 "<prompt>"                               # 4 images, auto-suffixed
ai-gateway image -m bfl/flux-2-pro -o cover.png "<prompt>"     # specific image-only model
ai-gateway image -m google/gemini-2.5-flash-image "<prompt>"   # Nano Banana (auto-routed via chat completions)

Video (saves .mp4 to disk, multi-minute job — prints a spinner with elapsed time):

ai-gateway video "<prompt>"                                            # default model, ./ai-video-<timestamp>.mp4
ai-gateway video -o clip.mp4 --duration 5 "<prompt>"                   # 5-second clip
ai-gateway video -m google/veo-3.1-fast-generate-001 \
  --aspect 9:16 --resolution 1080p -o vertical.mp4 "<prompt>"          # vertical 1080p Veo
ai-gateway video --json --duration 5 "<prompt>"                        # JSON with cost + elapsed

Success criteria:

• Text: streamed output ends with a newline, exit 0.
• Image: stdout contains Saved: <absolute path> and the file exists on disk.
• Video: stdout contains Saved: <absolute path> to a non-empty .mp4 (or .webm).

4. Report the result back

For text: relay the model output to the user (it's already on stdout). For image: report the absolute path(s) printed by the CLI. Do not re-encode or open the file unless asked.

Success criteria: User has the answer/file path.

Rules

• Never hardcode an API key in scripts. Resolution is env > config; agents should not read or write ~/.config/ai-gateway-cli/config.json directly.
• For multi-image batches with multimodal LLMs (Nano Banana etc.), the CLI loops n times against /v1/chat/completions since chat has no native batch — be patient with -n 4.
• Video generation is slow (often 1-5 minutes) and requires a $10 minimum balance on the AI Gateway account. Don't kill the spinner — wait it out.
• If the model id is unknown, run ai-gateway models --search <substring> to discover it instead of guessing.
• Don't pass --json to the image/video command if the user wants a human-readable result; it suppresses the friendly "Saved: ..." lines.
• The default text model is non-reasoning + fast; for harder reasoning tasks pick xai/grok-4.1-fast-reasoning, anthropic/claude-opus-4.6, or openai/gpt-5.4 via -m.

Troubleshooting

• Unauthorized (401) → key is wrong/expired. Reset with ai-gateway config set key <value>.
• Not found (404). Unknown model? → run ai-gateway models --search <hint>.
• Model "<id>" does not support image generation → the chosen model isn't an image-only model and lacks the image-generation tag. Pick from ai-gateway models --type image or use a multimodal LLM listed above.
• Video generation requires a minimum balance of $10 → top up at https://vercel.com/d?to=%2F%5Bteam%5D%2F%7E%2Fai%3Fmodal%3Dtop-up before retrying.