Levea MCP server

Levea — Agentic Video Production Platform

The professional agentic production environment for editable video. Levea turns natural-language creative direction, transcripts, and source media into fully structured, editable video projects, executes edits through a deterministic production harness, verifies the result, and renders delivery files.

Beta Notice: Agentic edits can make mistakes. Always preview every output before publishing. For production-safe and high-impact workflows, use requirePlanApproval: true to halt execution after the planning phase and inspect the proposed edit list.

⚡ The Levea Philosophy: Prompt-to-Project

Unlike traditional one-shot AI video generators that output locked, un-editable pixels, Levea maintains a fully structured, multi-layer project and timeline (Scene IR) containing assets, text, masks, audio, and brand kits.

We use probabilistic Frontier LLMs solely for planning, semantic analysis, and parameter parsing. The actual layout, timing, audio cleanup, face tracking, and composition are executed by a high-performance deterministic video-production harness. Generative media models are optional, modular assets; they do not own the project state.

🏗️ Platform Architecture

Levea separates probabilistic creative reasoning from deterministic project execution:

Creative Intent + Source Media
                                  │
                                  ▼
               Probabilistic Multimodal Intelligence
                          (Frontier Models)
                                  │
                                  ▼
                Typed Edit Graph / Media IR
                        (Scene Graph DAG)
                                  │
                                  ▼
               Deterministic Video-Production Harness
                   ├── Timeline and Scene Graph
                   ├── Media Operators
                   ├── Caption and Layout Engine
                   ├── Animation and Motion System
                   ├── Generative Media Adapters
                   ├── Composition and Asset Execution
                   │   ├── Remotion — cards, charts, diagrams and editorial compositions
                   │   ├── Vulkan/WebGPU — captions, primitives, effects and final compositing
                   │   ├── Lottie — verified authored vector assets
                   │   ├── HyperFrames — verified data-driven motion composites
                   │   └── Omni/Veo — generated supporting media
                   ├── Valid

1. Multimodal planning

The planning layer interprets user prompts, transcripts, visual references, and timeline boundaries. It compiles raw natural language into a highly optimized, typed Workflow DAG of sequential and parallel editing operations. The hosted planner currently utilizes Gemini through Google AI and Vertex AI.

Shortened. The full README is on GitHub.