Mov Studio AI Creation Desk
Chapter 01
Overview
A web-based AI film and television creation workbench. Creators work with the project Agent on a single canvas to complete scripts, character and scene assets, storyboards, image / video / voice-over generation, 3D previsualization and editing. Every AI call is first quoted in real time by the main site: generations are confirmed item by item, Agent inference is pre-deducted within the budget the user sets for the current round, and the final charge follows actual usage, with any excess refunded or any shortfall charged.
Chapter 02
Screenshots
01 image
-
01 Storyboard workspace: edit shot size, camera position and lines per shot, and generate previews shot by shot (requires logging in with a main-site account)
Chapter 03 05 items
My work
-
Migrated the project Agent’s execution core to the open-source OpenAI Codex CLI (embedded as an external process, with version-lock verification); connected the existing ~60 business tools (story, canvas, assets, generation, director’s desk, editing) to the Codex dynamic tool protocol, and added tool-receipt deduplication and interruption protection; tools that write data or incur charges must be confirmed by the creator.
-
Developed in-house a protocol gateway between Responses and Chat Completions, so the Agent can call DeepSeek / GLM / Kimi for inference through the main site; streams back the thinking content and automatically trims long sessions, meeting the requirements of multi-turn conversations with tools in DeepSeek’s thinking mode.
-
Duplicate-charge prevention: the request body at submission exactly matches the quote, and the user either confirms each charge or it is pre-deducted automatically within the current round’s budget; the idempotency key is written to disk first and the main site’s transaction number is recorded, and when a result is unknown the request is not automatically resent under a new number.
-
Agent permission narrowing: the Codex process uses a separate directory and minimal environment variables, has no access to main-site tokens or storage keys, has native shell, browser and similar capabilities all disabled, and can only call registered business functions.
-
Integrated the main site’s OIDC login and wallet APIs; user media is stored in Tencent Cloud COS (partitioned by account, read directly with short-lived signatures); took part in deploying and launching the web version.
Chapter 04 10 items
Tech stack
- React 18
- TypeScript
- React Flow
- three.js
- Tiptap
- Python
- FastAPI
- OpenAI Codex CLI
- FFmpeg
- Tencent Cloud COS
Chapter 05
Note
Sources and licenses
A team project; I was responsible for the parts listed above. The Agent core integrates the open-source Codex CLI (its source code is not modified); the editing desk is adapted from the open-source LTX-Desktop (Apache-2.0); the 3D director’s desk embeds the open-source Kunpeng Director.
