bibigpt-skill — Video Summarizer Agent Skill

Claude Code · Codex · ChatGPT · OpenClaw · Cursor. Summarize YouTube, Bilibili, podcasts, and local files via the bibi CLI or hosted MCP.

AI Agent skills for BibiGPT. Multiple skills, one CLI (bibi), one account.

Skill	When to use
`bibi`	Watch a video: summarize, transcript, chapters
`bibi-library`	Already-saved videos, notes, collections, in-library search
`bibi-feed`	Subscribe to channels, pull latest, mark seen
`bibi-vision`	Frames / OCR / slides, mind map

npx skills add JimmyLv/bibigpt-skill still installs bibi (existing onboarding). Add a sibling with --skill:

Terminal

npx skills add JimmyLv/bibigpt-skill --skill bibi-library
npx skills add JimmyLv/bibigpt-skill --skill bibi-feed
npx skills add JimmyLv/bibigpt-skill --skill bibi-vision

Agent landing: bibigpt.co/mcp. Hand-held per-host install pages: bibigpt.co/agent (ChatGPT, Claude Code, OpenClaw, WorkBuddy, Doubao Work, QwenWork, DeepSeek Harness).

Four ways to use:

BibiGPT Desktop + CLI Skill — install bibi CLI, works with Claude Code / OpenClaw / Codex / WorkBuddy / QwenWork / Doubao Work / DeepSeek Harness
Remote MCP Server — zero install, works with any MCP client (Claude, ChatGPT, Cursor, etc.)
OpenAPI — direct HTTP calls for containers, CI, or custom integrations
Codex Plugin — this repo doubles as a Codex plugin (.codex-plugin/ manifest + MCP + skills)

Install as a Codex plugin

From the terminal:

bash

codex plugin marketplace add JimmyLv/bibigpt-skill
codex plugin add bibi@bibigpt

Or inside a Codex session: /plugin marketplace add JimmyLv/bibigpt-skill, then install bibi from the marketplace tab in /plugins. Also works in the Codex IDE extension (Settings → Plugins) and the ChatGPT desktop app (Work / Codex). Authentication happens on first use via OAuth — no API keys to manage.

Skill Structure

Code

skills/
├── bibi/                 # Watch: summarize / subtitle / chapter
├── bibi-library/         # Saved videos, notes, collections
├── bibi-feed/            # Channel subscribe + latest feed
└── bibi-vision/          # Frames, OCR, mind map

Each directory has SKILL.md (intent + triggers), references/ (auth, CLI), and workflows/ where a multi-step path exists. All four skills call the same bibi binary — there is no second npm package named video-cli.

Code

skills/bibi/
├── SKILL.md                        # Watch-video router — library/feed/vision go to siblings
├── scripts/
│   └── bibi-check.sh              # Auto-detect CLI vs API mode
├── references/
│   ├── cli.md                     # CLI command reference
│   ├── api.md                     # OpenAPI endpoint reference
│   ├── endpoints.md               # Auto-generated OpenAPI dump
│   ├── installation.md            # Setup & auth guide
│   └── supported-platforms.md     # URL types & platform limits
└── workflows/
    ├── quick-summary.md           # Paste URL → get AI summary
    ├── deep-dive.md               # Chapter breakdown + follow-up Q&A
    ├── transcript-extract.md      # Subtitle/transcript extraction
    ├── article-rewrite.md         # Video → blog/公众号图文/小红书
    ├── batch-process.md           # Multi-URL batch processing
    ├── research-compile.md        # Multi-source topic synthesis
    ├── export-notes.md            # Save to Notion/Obsidian/local
    └── visual-analysis.md         # Kept as fallback; prefer skill `bibi-vision`

Workflows

bibi (watch-video) workflows:

Workflow	What it does	Trigger examples
Quick Summary	One URL → AI summary	"summarize this video", "总结"
Deep Dive	Chapter-by-chapter + Q&A	"chapter summary", "分章节总结"
Transcript Extract	Raw subtitles with timestamps	"get subtitles", "获取字幕"
Article Rewrite	Video → polished article	"turn into article", "AI改写"
Batch Process	Multiple URLs at once	"batch summarize", "批量总结"
Research Compile	Cross-source synthesis	"compare these videos", "综合分析"
Export Notes	Save to Notion/Obsidian/file	"save to Notion", "导出笔记"

Sibling skills own the rest: bibi-library (saved library / notes / collections), bibi-feed (channels / latest), bibi-vision (frames / mind map).

Quick Start

1. Install Desktop App

Terminal

curl -fsSL https://bibigpt.co/install.sh | bash

Works on macOS (auto-uses Homebrew), Linux (downloads AppImage), and detects Windows (prompts to use winget).

Or install manually per platform:

macOS: brew install --cask jimmylv/bibigpt/bibigpt (Homebrew tap)
Windows: winget install BibiGPT --source winget or download from bibigpt.co/download/desktop
Linux: Download .deb / .AppImage from bibigpt.co/download/desktop

Then log in: bibi auth login (opens browser for OAuth, saves token automatically).

2. Install Skill

Claude Code

Terminal

npx skills add JimmyLv/bibigpt-skill

OpenClaw

Terminal

npx skills add JimmyLv/bibigpt-skill --agents OpenClaw --yes

DeepSeek Harness

bash

dsh plugin --profile web add "github:JimmyLv/bibigpt-skill#path:/dsh-plugin"

Restart the profile and /bibi shows up under Skills. Keep the quotes — # starts a shell comment. Source and the disk-copy alternative live in dsh-plugin/. Copying skills/bibi into ~/.agents/skills/bibi also works (shared with Claude Code).

WorkBuddy (Tencent)

Download: workbuddy.cn. Skills live in ~/.workbuddy/skills/<name>/SKILL.md.

Terminal

npx skills add JimmyLv/bibigpt-skill -g -y

WorkBuddy custom skills need agent_created: true in the SKILL.md frontmatter (this repo includes it). Then /reload-skills or restart the app.

Fallback: copy skills/bibi to ~/.workbuddy/skills/bibi.

Step-by-step: https://bibigpt.co/agent/workbuddy

豆包工作 / Doubao Work

Download desktop: doubao.com/download/desktop. Doubao Work does not publish a filesystem skill directory — do not invent ~/.doubao/skills.

In the desktop sidebar open 技能 · 连接器 · 伙伴, create a skill named bibi, and paste skills/bibi/SKILL.md.

Step-by-step: https://bibigpt.co/agent/doubao

千问工作 / QwenWork

App: qwenwork.cn. Skill dir: ~/.qwenworkcn/skills/. Paste this into QwenWork:

text

Please download https://github.com/JimmyLv/bibigpt-skill and copy skills/bibi to ~/.qwenworkcn/skills/bibi

QwenWork clones and loads the skill from chat. Docs: https://help.aliyun.com/zh/qwenwork/skills

Step-by-step: https://bibigpt.co/agent/qwenwork

3. Usage

Ask your agent to summarize any video, audio URL, or local file:

Code

> Summarize this video: https://www.youtube.com/watch?v=xxxxx
> Summarize this local file: /path/to/meeting-recording.mp4

The agent will automatically detect the best mode and route to the right workflow.

Commands

Command	Description
`bibi summarize "<URL>"`	Summarize a video/audio URL
`bibi summarize "/path/to/file.mp4"`	Summarize a local audio/video file
`bibi summarize "<INPUT>" --async`	Async mode (long videos)
`bibi summarize "<INPUT>" --chapter`	Chapter-by-chapter summary
`bibi summarize "<INPUT>" --subtitle`	Fetch subtitles/transcript only
`bibi summarize "<INPUT>" --json`	Full JSON output
`bibi auth check`	Check auth status
`bibi auth login`	OAuth login via browser (saves token automatically)
`bibi check-update`	Check for new version
`bibi upgrade`	Download and install latest version (alias: `self-update`)

Updating

BibiGPT ships three independently updateable layers. Pick the one that matches what you want refreshed.

Layer	Updates	Command	When to run
Skill content (workflows, references, intent router)	npm-distributed, manual	`npx skills update JimmyLv/bibigpt-skill`	Weekly, or when changelog adds a new MCP tool / workflow
`bibi` desktop CLI (binary + embedded SKILL.md + manifest dispatcher)	Self-update or package manager	`bibi upgrade` (or `brew upgrade --cask bibigpt` / `winget upgrade BibiGPT`)	When the changelog mentions CLI changes
Remote MCP server (`https://bibigpt.co/api/mcp`)	Always latest, server-side	automatic — restart your MCP client to pick up new tools	Never manually; happens server-side

Update the skill

Terminal

npx skills update JimmyLv/bibigpt-skill

The skill does not auto-update — Agents read whatever was last fetched. Re-run when you want new BibiGPT capabilities, fixed workflows, or new MCP tools surfaced. For a clean reinstall:

Terminal

npx skills remove JimmyLv/bibigpt-skill && npx skills add JimmyLv/bibigpt-skill

The GitHub Releases page tracks what each version changed.

Update the bibi CLI

bash

bibi check-update    # peek at the latest published version
bibi upgrade         # download and install (uses brew on macOS, installer on Windows)

Or via package manager: brew upgrade --cask bibigpt / winget upgrade BibiGPT --source winget / fresh download from bibigpt.co/download/desktop.

Manifest is hot-cached, not embedded. Even without bibi upgrade, the CLI dispatcher fetches https://bibigpt.co/api/cli-manifest.json on first call (24h TTL) — newly-released agent procedures show up the next day. bibi upgrade only matters when the dispatcher itself or the embedded SKILL.md changes.

Update the MCP server

Nothing to do — the remote MCP endpoint always serves the latest meta.mcp.enabled tool list. Restart your MCP client (Claude Desktop, Cursor, etc.) if you don't see a newly-released tool — that refreshes its tools/list cache.

Note: The desktop app build is triggered via the build-tauri-desktop-app GitHub Actions workflow. After a successful build, artifacts are uploaded to OSS and the latest.json manifest is updated automatically — the client checks this manifest for update availability.

Remote MCP Server (No Install Required)

BibiGPT also provides a remote MCP server at https://bibigpt.co/api/mcp — works with any MCP-compatible client. Streamable HTTP transport with OAuth 2.1 authentication.

Available Tools

Tool	Description
`summarize_video`	Summarize a video or podcast URL
`summarize_video_with_config`	Summarize with custom prompt, model, language
`summarize_by_chapter`	Chapter-by-chapter summary
`get_subtitle`	Get transcript/subtitles with timestamps
`create_summary_task`	Create async task for long videos
`get_task_status`	Check async task status

Claude Code

Add to .mcp.json in your project root:

config.json

{
  "mcpServers": {
    "bibigpt": {
      "type": "http",
      "url": "https://bibigpt.co/api/mcp"
    }
  }
}

Or via CLI:

Terminal

claude mcp add --transport http bibigpt https://bibigpt.co/api/mcp

OpenAI Codex

Edit ~/.codex/config.toml (or .codex/config.toml in project root):

toml

[mcp_servers.bibigpt]
url = "https://bibigpt.co/api/mcp"

With API key authentication:

toml

[mcp_servers.bibigpt]
url = "https://bibigpt.co/api/mcp"
bearer_token_env_var = "BIBIGPT_API_KEY"

Claude Desktop

Option A — UI (recommended): Settings → Connectors → Add Connector → paste https://bibigpt.co/api/mcp

Option B — Config file (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

config.json

{
  "mcpServers": {
    "bibigpt": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://bibigpt.co/api/mcp"]
    }
  }
}

Cursor

Add to .cursor/mcp.json in project root, or configure in Cursor Settings → MCP:

config.json

{
  "mcpServers": {
    "bibigpt": {
      "url": "https://bibigpt.co/api/mcp",
      "type": "streamable-http"
    }
  }
}

VS Code (GitHub Copilot)

Add .vscode/mcp.json in your workspace:

config.json

{
  "servers": {
    "bibigpt": {
      "type": "http",
      "url": "https://bibigpt.co/api/mcp"
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

config.json

{
  "mcpServers": {
    "bibigpt": {
      "serverUrl": "https://bibigpt.co/api/mcp"
    }
  }
}

ChatGPT

Settings → Apps & Connectors → Advanced settings → enable Developer mode
Settings → Apps & Connectors → Create connector
Enter URL: https://bibigpt.co/api/mcp
Complete OAuth authorization when prompted

Manus

Settings → Connectors → Add Connectors → Custom MCP → +Add Custom MCP:

Enter server URL: https://bibigpt.co/api/mcp

LobeChat

Settings → Plugin Settings → Custom Plugins → Quick Import JSON:

config.json

{
  "type": "mcp:streamable-http",
  "url": "https://bibigpt.co/api/mcp",
  "metadata": {
    "title": "BibiGPT",
    "description": "AI Video & Audio Summarizer"
  }
}

Direct API Key (Skip OAuth)

If you have a BibiGPT API key, skip OAuth and use Bearer token directly:

bash

# List available tools (no auth needed for discovery)
curl -X POST https://bibigpt.co/api/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# Summarize a video (requires auth)
curl -X POST https://bibigpt.co/api/mcp \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"summarize_video","arguments":{"url":"https://www.youtube.com/watch?v=xxxxx"}}}'

Get your API key at bibigpt.co/user/integration.

License

MIT