Media generation

Image tools for the agent: generate_image, edit_image, and an experimental generate_video. Read the limits first — they're the point.

Honest limits

• Image generation is never free-unlimited. Two real providers:
  1. OpenAI Images API (gpt-image-1) — a paid API. Every call spends your money on your key. The tools are medium risk, so the permission gate asks before each call.
  2. Local Stable Diffusion — the AUTOMATIC1111 web UI API, if you run it yourself. Free after your own hardware.
• No provider, no image. Without a key or a reachable local server, the tools return a configuration error. Nothing is faked.
• Video is experimental. AgentKai ships no bundled video provider; generate_video returns an explanatory error until you register one (see VideoProvider in agentkai/media.py).

Setup

Option A — OpenAI (paid):

1. Get a key at https://platform.openai.com/api-keys (billing enabled — image calls are charged per image).
2. export OPENAI_API_KEY=..., or create ~/.agentkai/openai_token.json:

   { "access_token": "<your-key>" }

Option B — local Stable Diffusion (free after hardware):

1. Install the AUTOMATIC1111 Stable Diffusion web UI, launch with --api (default http://127.0.0.1:7860).
2. No key needed. Set AGENTKAI_IMAGE_PROVIDER=local_sd to force it.

Provider selection order: config["media"]["image_provider"] → AGENTKAI_IMAGE_PROVIDER env → auto (OpenAI key present → OpenAI; local SD answering → local SD; else an honest error).

Editing

edit_image does inpainting/edits (OpenAI provider; mask optional). Generated files land in ~/.agentkai/media/ with unique names.

What's not here

• No bundled video generation. The generate_video entry point exists so a future provider can plug in; today it tells you exactly that.
• No face/person identification — that's a hard no, not a missing feature.