Case study · MCP platform
studio-mcp
An MCP server that turns a full film pipeline — brief → plan → render → QC → cut — into 14 agent-callable tools. An LLM drives the generation stack directly; a human approves at the gates.

MARÉA shot 1 · Higgsfield Soul Cinema · 21:9 · passed the vision QC gate at intent 80 / look 84 / character 90
The grading bay — live console
studio·mcp — grading bay14 instruments · ● live
- 00the director
- 01plan_shots
- 02lock_campaign
- 03qc_still
- 04assemble
- 05gen_still
- 06animate
- 07train_character
- 08project_status
- 09reference_prompt
- 10palette_from_image
- 11cut
- 12upscale
- 13craft_lookup
- 14list_models
instrument · qc_still
Score a render before it ships.
The bay speaks to the pipeline over studio-mcp --transport streamable-http. Every take is timed, priced, and logged to the reel — schema-driven forms, streamed renders, an LLM Director persona operating the tools.
How it works
- Orchestration graph — Planner → Generation ⇄ QC → Assembly — with retries, human gates, and replayable per-run traces.
- 14 tools in four groups: plan & lock · render & QC · animate & train · assemble & ground.
- Vision QC gate:
qc_stillscores every render against shot intent and the locked look, returning pass/fail plus a fix suggestion — style drift gets caught, not shipped. craft_lookupgrounds prompts in creative-rag, a cited film-craft knowledge base.- Web console — a "grading bay" with schema-driven forms, streamed renders, and an LLM Director persona that operates the tools. Live on Vercel.
- Provider-agnostic LLM layer (any OpenAI-compatible endpoint) · state is plain JSON on disk · Langfuse tracing + an offline benchmark suite.
Details
14MCP tools, four functional groups
3.10–3.12Python versions in CI matrix, green
3–5reference photos to train a Soul-ID self-clone
2animate modes — DP-persona direct, or full timecoded Seedance prompt
Honest note: craft_lookup needs creative-rag running alongside — the two ship as a pair.