Selected work / Case study
codex-video-vision
Video tools for Codex. A credited port of an MIT-licensed project.
Tests checkedOverview
A Codex video tool server using MCP, TypeScript, and ffmpeg.
- TypeScript
- MCP
- ffmpeg
The problem
A coding agent needs timestamped frames and audio information to inspect a local video. This project prepares metadata, images, scene analysis, and optional transcripts for the agent to interpret.
My contribution
I adapted the MIT-licensed claude-video-vision project for Codex, preserving its upstream credit and license. The port includes a local MCP server and workflows for video inspection and storyboard comparison.
How it's built
A TypeScript server exposes MCP tools for metadata, analysis, frame extraction, and closer inspection of selected timestamps. ffmpeg and ffprobe process local media. A session cache supports later frame retrieval. Transcription can use Gemini, OpenAI Whisper, or a local Whisper backend.
What's verified
108 tests pass.
Eyrie reported 108 passing tests on October 29, 2026. This is a credited port of an MIT-licensed project. No live demo or end-to-end transcription check is claimed.
What remains unverified
Transcription tests use mocked backends; real API transcription and local Whisper inference were not verified. Codex Desktop plugin discovery and long-video performance also remain unchecked. The report validates the local source build, not a published npm installation or a hosted demo.
Captured screens
