klo-mcp
Enables AI agents to control the klo macOS video editor locally, providing 83 tools for timeline editing, captions, color grading, and export while keeping media on-device.
README
<p align="center"> <a href="https://getklo.com"><img src="https://getklo.com/og.jpg?v=3" alt="klo, the native Mac video editor you talk to" width="720" /></a> </p>
klo-mcp: the video editing MCP server
Give your coding agent a real video editor.
klo is a native macOS video editor you talk to. This is its MCP server: Claude Code, Codex, Cursor, Claude Desktop, or any MCP client can drive the whole timeline, with 83 tools, while your footage never leaves your Mac.
- A real editor underneath. Not an ffmpeg wrapper: ripple edits, word-gap cuts, karaoke captions, color grades, generated music, ProRes export. Every edit lands on a normal timeline a human can open and adjust.
- Media stays local. Piping video bytes through a text protocol fails (base64 versus context windows). klo holds the media on your machine and exposes semantic timeline operations instead. The agent sends intent and gets back compact results: timecodes, transcripts, real frames when it needs to look.
- Local, token-gated, off by default. The server listens on
127.0.0.1only, requires a bearer token, and does nothing until you turn it on in klo's settings.
Quickstart
- Get klo for Mac (free tier included, macOS 13+).
- In klo: Settings → Connections, turn the connection on, and copy the URL and token it shows.
- Connect your client:
Claude Desktop (one click)
Download klo.mcpb and open it. Done.
Claude Code
claude mcp add --transport http klo http://127.0.0.1:41720/mcp \
--header "Authorization: Bearer YOUR_TOKEN"
Use the exact URL shown in klo's settings (the port can differ if 41720 is taken).
Cursor, Codex, and other MCP clients
Either point the client at the HTTP endpoint above, or use this package as a stdio server:
{
"mcpServers": {
"klo": {
"command": "npx",
"args": ["-y", "klo-mcp"],
"env": { "KLO_MCP_TOKEN": "YOUR_TOKEN" }
}
}
}
What your agent can do
Ask in plain language, or call tools directly. A few real prompts:
- "Cut the boring parts and the filler words, then caption it word by word."
- "Watch this folder and turn every new recording into a captioned vertical clip."
- "Pick the best take of each line from the transcript and assemble the cut."
- "Grade everything to match this reference and export ProRes for the client."
The 83 tools
| group | tools |
|---|---|
| library | get_project_state, get_media_library, import_media, import_from_url, relink_to_sources, list_files, new_project |
| timeline | add_media_to_timeline, assemble_cut, insert_at, move_clip, split_clip, split_audio, trim_clip, delete_clips, ripple_delete, close_gaps, separate_audio, unlink_clip, link_clips, replace_audio, add_marker, set_in_out, set_track_lock, seek, verify_timeline |
| project | set_canvas_ratio, set_frame_rate, set_brand_kit |
| cuts | add_transition, remove_transition, preview_cut, generate_transition, refine_transition |
| type | add_text, update_text, add_captions_from_transcript, list_text_styles, list_fonts, search_fonts, use_font |
| motion | list_motion_templates, create_motion_clip, update_motion_clip |
| color | set_clip_grade, add_clip_effect, update_clip_effect, remove_clip_effect, list_luts, import_lut, set_clip_opacity, set_clip_blend |
| frame | set_clip_transform, set_clip_speed |
| audio | set_clip_audio, set_volume_envelope, duck_audio, analyze_audio, quantize_cuts_to_beats, isolate_voice_on_clip, measure_audio, measure_mix |
| generate | generate_music, generate_sfx, generate_voiceover, clone_voice_from_clip, fix_spoken_line, generate_video |
| read | get_transcript, find_repeated_takes, capture_frame, survey_footage, measure_frame, measure_clip |
| remember | write_footage_notes, get_footage_notes, analyze_reference, save_style_profile, list_style_profiles, apply_style_profile |
| out | delegate_research, web_search, export_video |
Pricing
Driving the timeline over MCP is free: your agent brings its own model. The handful of tools that generate new media (music, voiceover, generated shots) use your klo account's credits, and the agent is told the price before it spends. New accounts get 10 free asks.
Security notes
- Listens on
127.0.0.1only. Nothing is exposed to the network. - Bearer token auth with constant-time comparison. Rotate the token from klo's settings anytime.
- Off by default. If klo is running but the connection is off, the server tells your agent exactly that.
- Voice cloning tools are consent-gated in the app.
Troubleshooting
"klo is running but is not accepting connections": turn the bridge on in
klo → Settings → Connections, then copy the token into KLO_MCP_TOKEN (or
the Authorization header for HTTP).
Port differs from 41720: klo falls back to an ephemeral port if the preferred one is taken. Always trust the URL shown in Settings.
About
This package is the thin connector; the editor itself is klo. The connector mirrors the copy bundled inside the app, and its major.minor version tracks the app release it shipped with.
FAQ
Can Claude Code edit videos? With klo installed, yes. Claude Code connects to this server and gains cutting, captions, grading, generated music and voiceover, and ProRes export on a real timeline. See what agentic video editing is.
Is this an ffmpeg wrapper? No. Tools operate on klo's timeline (ripple edits, word-gap cuts, undo), and nothing is re-encoded between steps. Why that design.
Does my footage get uploaded? No. Media stays on the Mac; the server exchanges timeline operations and compact results. Details in klo's privacy policy.
What does it cost? Driving the timeline is free with your own model. Generation tools use klo account credits. Pricing.
Related reading
- Give your coding agent a real timeline
- What is agentic video editing?
- klo changelog: the app ships several times a week
- Download klo for Mac
MIT (this connector). klo is a product of klorah inc. getklo.com
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。