All plugins

rea-video-hermes

Reverse-engineer any video shot by shot: production-method classification per shot (live action / AI video / 3D / screen capture / code-template render), caption and sticker timelines, voice vs music stems, editing rhythm, and a concrete replication plan. Powered by JuneYaooo/rea-video (Apache-2.0, pinned commit d904b1e8). The six stdio-MCP tools expose the full pipeline (incl. mux_tracks for dual-track video downloads); for shots the pipeline flags as low-confidence, the agent reviews the tiled frames with its own vision and merges the verdict back via submit_review — no external VLM key required. Local runtime deps: ffmpeg and the pinned rea-video package (installed from the repo's venv). Disclosure — downloads video/audio from Douyin share and CDN URLs the agent passes in (https only, Douyin hosts only, curl with --noproxy), runs the third-party rea-video package (git-pinned) and ffmpeg on the host, writes all outputs under its own plugin-data folder (falling back to ~/rea-video-out), and the upstream pipeline downloads a ~67 MB audio-separation model on first run.

wusanjayToolsv0.1.0Updated 5 days ago

About

Reverse-engineer any video shot by shot: production-method classification per shot (live action / AI video / 3D / screen capture / code-template render), caption and sticker timelines, voice vs music stems, editing rhythm, and a concrete replication plan. Powered by JuneYaooo/rea-video (Apache-2.0, pinned commit d904b1e8). The six stdio-MCP tools expose the full pipeline (incl. mux_tracks for dual-track video downloads); for shots the pipeline flags as low-confidence, the agent reviews the tiled frames with its own vision and merges the verdict back via submit_review — no external VLM key required. Local runtime deps: ffmpeg and the pinned rea-video package (installed from the repo's venv). Disclosure — downloads video/audio from Douyin share and CDN URLs the agent passes in (https only, Douyin hosts only, curl with --noproxy), runs the third-party rea-video package (git-pinned) and ffmpeg on the host, writes all outputs under its own plugin-data folder (falling back to ~/rea-video-out), and the upstream pipeline downloads a ~67 MB audio-separation model on first run.

hermes-plugin-rea-video

给 Hermes Agent 装上视频逆向能力:丢一条视频(或抖音链接),Hermes 逐镜告诉你
实拍 / AI 生成 / 3D / 录屏各占多少、字幕花字的时间轴、声音是不是 AI、以及怎么复刻一条一样的。

Demo:让 Hermes 当面拆解一条视频 → 逐镜判定 → 对低置信镜头亲自看图复核。

tool 51% AI vs agent review 97% real — 融合复核证据

端到端实测(2026-11-11):63 秒真实视频走完整循环——16 镜分解(AI 42% / 模板渲染 39% / 静图 13%),agent 复核最争议镜头(ai_video 0.45 vs static_hold 0.29)后判 code_render 并回写融合。证据:demo/e2e_run/(report.md · breakdown.json · storyboard · agent transcript)。

Powered by JuneYaooo/rea-video (Apache-2.0) —
本仓库只是 Hermes 插件封装(MCP server + skill),全部核心能力来自上游工具,依赖钉死在
commit d904b1e8328515dc95d9772cd2e65fb8acde996d(Apache-2.0 第 9 条不可撤回,上游任何后续变化不影响本插件)。
上游贡献:shot 复核的 split_suggestion 与 disputed-review 融合规则由本仓库维护者的 PR 已合入上游。

架构

用户: "拆一下这条视频" + 链接
  └─ resolve_douyin   分享链接 → 视频页(下载由用户/工具完成)
  └─ decompose_video  15 镜分类 + 142 句字幕时间轴 + 人声/配乐分离
  └─ video_summary    结论 + 低置信镜头清单
  └─ shot_tiles × N   Hermes 用自己的视觉能力看图复核(agent-as-VLM)
  └─ submit_review    复核并入,报告定稿
→ report.html / report.md / subtitles.srt / 复刻方案

安装

pip install -e .
# 运行需要 ffmpeg + Python 3.10+

Hermes 注册:把 mcp_server.py 以 stdio MCP server 挂进 Hermes 配置,skills/ 放入技能目录。

提交到官方插件目录

参见 catalog/hermes-plugin-rea-video.yaml(对 hermes-agent 仓库 plugin-catalog 的 PR 即收录)。

Demo

见 demo/DEMO.md(脚本与分镜)。

README from the reviewed commit 5d2bc03; it updates when the author re-pins.

Developer Mode

Explore built-in Hermes toolsets and actions.

HermesAgent

Developer?

Create a Hermes extension for Desktop or agent workflows.

Developer Section