dsh-plugin-chatcut

Edit videos with AI, right inside DeepSeek Harness.

Bring ChatCut — the agent-native video editor — to DSH (DeepSeek Harness) as a first-class MCP plugin: 50 editing tools + 15 craft skills, with one-time OAuth setup and fully automatic token refresh.

中文文档 · Quick start · How it works

#dsh-plugin #mcp #video-editing #chatcut

What you get

Ask your DSH agent, in plain language:

• 🎙️ Talking-head cleanup — "remove all filler words and awkward pauses" (transcript-based editing: the agent edits text, the timeline follows)

• ✂️ Highlight cuts — "make a 30s version about pricing", "keep the best take"

• 💬 Captions — styled, bilingual, word-level karaoke highlight

• 🎨 Motion graphics — generated from JSX code, or from ChatCut's template library

• 🔍 Zoom, transitions, LUTs, sound effects — from the built-in library

• 🖼️ AI generation — images (GPT Image), video (Seedance), music, TTS, voice clone

• 📤 Export — mp4 / audio / SRT / NLE XML (Premiere · DaVinci · FCP) / transparent-background MG ProRes

Everything lands on a real multi-track NLE timeline you can also edit by hand at chatcut.io — the agent and you share the same project, live.

Why this plugin

ChatCut officially ships plugins for Claude Code and Codex only. This project adapts it to DSH:

Configure the token once. Never think about it again. The bridge refreshes the access token with the rotating refresh token and stores it locally (0600).

Quick start

Prereqs: a ChatCut account (free tier works), DSH Desktop, network access to api.chatcut.io.

git clone https://github.com/nigedazhima/dsh-plugin-chatcut.git
cd dsh-plugin-chatcut

1. Log in (one time) — opens your browser for ChatCut OAuth:

# macOS with DSH Desktop (no separate Node needed):
ELECTRON_RUN_AS_NODE=1 "/Applications/DSH Desktop.app/Contents/MacOS/DSH Desktop" scripts/chatcut-login.mjs
# or, with Node >= 18:
node scripts/chatcut-login.mjs

Tokens are written to scripts/.tokens.json (git-ignored, mode 0600).

2. Install the skills:

cp -R skills/chatcut "$HOME/Library/Application Support/dsh-desktop/harness/skills/"

3. Mount the MCP server — copy cordis-patch-example.yml into your $DSH_HOME/cordis.patch.yml, replacing <REPO> with this repo's absolute path.

4. Restart DSH Desktop. In a new session, ask: "list your mcp__chatcut__ tools" — you should see ~50 tools.

How it works

DSH agent ──(stdio MCP)── chatcut-bridge.mjs ──(streamable-http + Bearer)── api.chatcut.io
                              │
                              ├─ 401? → refresh_token → new access_token → persist
                              └─ .tokens.json (0600, git-ignored)

• The bridge reuses the official @modelcontextprotocol/sdk already bundled with DSH — zero npm installs on macOS.

• Skills are the official ChatCut craft skills (talking-head guide, motion graphics, captions, export, …) plus a thin DSH host-adapter that maps Claude-Code-isms (tool prefixes, login commands, browser panes) to DSH equivalents.

• Transcript-first editing: the killer workflow is read_script → edit a markdown file with strikethrough → apply_script. Deleting a sentence in text deletes it on the timeline.

Troubleshooting

Security notes

• No credentials ship with this repo. .tokens.json is created locally by you and git-ignored.

• The OAuth client is registered dynamically (RFC 7591); no shared client secrets.

• Token file is chmod 0600; refresh tokens rotate on every refresh.

License

GPL-3.0-only — same as the upstream ChatCut agent-plugin, from which the skills/chatcut/resources/ are redistributed unchanged.

Credits

• ChatCut for the editor and the official agent plugin

• DeepSeek Harness for the pluggable agent runtime

If this saved you an editing afternoon, a ⭐ helps others find it.