bench-studio

Bench Studio

Use Bench as a local creative operating system. Prefer its MCP tools; fall back to the loopback API at http://localhost:8787 when the MCP server is not connected.

Route the request

• Images or videos: inspect models, inspect the chosen model's inputs, upload any media, then generate.
• Websites: start create_website, retain the project id, poll get_project, and return the preview plus source artifact.
• Documents: start create_document, poll get_project, and return the PDF plus editable HTML preview.
• Existing work: use list_results for media or list_projects for websites and documents.
• Pricing or storage: use get_usage before making assumptions.

Media workflow

1. Call list_models with the requested output and required input modality.
2. Call get_model_capabilities before attaching media. Treat schema-supported as declared support, not proof of visual fidelity.
3. Call upload_media for every local asset. Keep both its hosted URL and local archive URL.
4. Map each uploaded asset to an exact capability field. Never invent image_url; models differ.
5. Call create_media with the smallest suitable model and explicit parameters.
6. Return local result URLs first and hosted fal URLs second.

Do not say a reference influenced an output merely because fal accepted the field. Say “submitted successfully” unless an output was actually reviewed.

Website and document workflow

Write a concrete creative brief containing audience, purpose, mood, required sections, and constraints. Pick low depth for routine drafts and medium for consequential work. Builds use the local ChatGPT-authenticated Codex SDK and create inspectable files under Bench's project archive.

Poll until complete, failed, or cancelled. Do not report a queued job as delivered. For PDFs, visually inspect representative rendered pages before calling the document finished when the user asks for a final production artifact.

Safety and truthfulness