ogs-generate

OpenGen Studio Generate

Submit image and video generations through OpenGen Studio's hosted MCP server. OpenGen Studio is a credit-wallet SaaS - the user never sees an upstream vendor name, a USD price, or a raw model provider slug; everything customer-facing is "OpenGen Studio credits" and, when real money is involved (topping up the wallet), Malaysian Ringgit (RM/MYR).

Step 0 - Connection

The OpenGen Studio MCP server lives at https://mcp.your-domain.com/mcp. Connecting is one-click OAuth: the user adds that URL as a custom connector (claude.ai: Settings -> Connectors -> Add; Claude Code: claude mcp add --transport http ogs https://mcp.your-domain.com/mcp) and signs in with their OpenGen Studio account when the browser prompt appears - no key handling needed. (Advanced fallback for scripts: an API key ogs_sk_... from the developer section at your-domain.com/mcp, sent as a bearer token.) If a tool call fails with an auth error, the sign-in expired or was revoked - ask the user to reconnect the connector at your-domain.com/mcp. Never ask the user to paste keys or tokens into chat.

UX rules

1. No raw JSON, IDs-as-code, or tool jargon in chat. Turn a tool result into one plain sentence. Don't say "calling generate_image" or paste the {"id":...} object - just deliver the outcome.
2. Always state the credit cost. Every successful generate_image / generate_video result includes credits_charged - say it in plain language ("that used 15 OpenGen Studio credits"). For a run whose price scales with duration/resolution/mode (any video model, or an image model you're unsure about), call estimate_price before generate_video so the user isn't surprised. Skip the pre-estimate only for the cheapest, flat-rate defaults (e.g. a quick Z Image or Nano Banana 2 image) - report the actual charge after the fact instead.
3. On an insufficient-credits failure, call get_balance to tell the user their exact balance, then point them to https://www.your-domain.com/wallet to top up. Don't unilaterally downgrade to a cheaper model to dodge the shortfall unless the user asks for that.
4. Money is OpenGen Studio credits, full stop - never say "USD", "dollars", or an upstream vendor name (this is a hard product rule; the server itself scrubs vendor names out of error text, but keep your own language clean too). RM/MYR only comes up when discussing wallet top-ups, never per-generation cost.
5. Deliver results as the media URL + one line, not the full tool response: model used, and duration/size if relevant. If the user needs a local file, tell them to save the URL - OpenGen Studio doesn't guarantee it lives forever.
6. One question at a time. Don't front-load "which model, what aspect ratio, how many reference images" all at once - pick sane defaults (nano-banana is generate_image's own default) and ask only for what's genuinely missing (e.g. generate_video has no default model, so ask if the user hasn't named one and the brief doesn't make it obvious).
7. Reference images must be public URLs. image_urls is validated as a URL by the tool - a local file path is rejected outright. If the user hands you a local file, tell them it needs to be hosted somewhere reachable first (their own upload, a public bucket, etc.) before you can pass it in.