26 MCP tools that give Claude Code, Cursor, and any AI agent the ability to trim, merge, overlay text, sync audio, apply filters, and export video. Local, fast, free.
AI agents can write code, analyze documents, browse the web, create images — but video editing? Existing tools are either GUI-only (agents can't click buttons), raw FFmpeg wrappers (agents can't memorize hundreds of flags), or cloud APIs (expensive, slow, vendor lock-in). mcp-video bridges this gap.
Trim, merge, add text, sync audio, resize, convert, speed change, watermarks, subtitles, storyboards, crop, rotate, fade, filters, color grading, audio normalization, PiP, split-screen, batch processing, and more.
Real-time progress reporting on long operations. Agents can tell you "50% done..." instead of going silent during converts and merges.
Runs on your machine. No cloud uploads, no API keys, no per-minute billing. Your video never leaves your computer.
Returns a thumbnail of the output after every operation. Agents can confirm results look right without opening the file.
Parses FFmpeg errors into structured responses with actionable suggestions. "Codec error: vp9" becomes "Auto-convert from vp9 to H.264/AAC."
Blur, sharpen, brightness, contrast, saturation, grayscale, sepia, invert, vignette. Plus color presets: warm, cool, vintage, cinematic, noir.
Normalize loudness to LUFS targets: YouTube (-16), broadcast (-23), Spotify (-14). Correct true peak for broadcast compliance.
Picture-in-picture overlay with positioning, opacity, and timing. Side-by-side or top/bottom split-screen with auto-resize.
Apply the same operation to multiple files in one call. Trim 10 videos, blur 5, or convert a whole directory at once.
Filter types, color presets, and parameters are validated before hitting FFmpeg. Get clear error messages, not cryptic KeyError tracebacks.
"Can't I just have Claude Code run ffmpeg commands?" Sure — if you enjoy debugging 200-character filter graphs, wrestling with flag order semantics, and paying dollars per short clip on cloud APIs. Here's why mcp-video is different.
| mcp-video | Claude Code + FFmpeg | Cloud APIs | |
|---|---|---|---|
| Cost | Free, local | Free, but your time isn't | $0.28–$2.50 per 5s video |
| Privacy | 100% local | 100% local | Uploads to cloud |
| Error messages | Structured + auto-fix | FFmpeg stderr (cryptic) | HTTP errors |
| Parameter validation | Before execution | Only at runtime | API schema check |
| Text overlay | add_text(text="Hi") |
-vf "drawtext=text='Hi':fontfile=...:fontsize=24:x=(w-text_w)/2:y=h-50" |
Varies by API |
| Speed change | Handles audio-less videos | Fails if no audio stream | Usually handled |
| Discovery | 26 self-documenting tools | Must know hundreds of flags | API docs |
| Testing | 380 tests, CI-ready | None (ad-hoc) | Provider's tests |
| Progress reporting | Real-time % | Silence until done | Varies |
| Visual verification | Thumbnail after ops | No | Varies |
Put -ss 10 before -i input.mp4 and FFmpeg seeks instantly. Put it after, and it decodes every frame up to 10s first. Same flags, completely different performance. Claude Code can't know which you meant.
Try adding audio to a silent video and you get: "Stream specifier ':a' matches no streams." No hint that the video has no audio. mcp-video checks before running and gives you a clear error.
Every FFmpeg call re-encodes. Chain 5 operations and you've re-encoded 5 times. mcp-video's timeline DSL composes operations into a single FFmpeg command where possible.
In a 10-step workflow, Claude must re-discover file paths, resolutions, and durations at every step. mcp-video returns structured results (path, duration, resolution) so the agent always has context.
Replicate charges $0.28–$0.50 per second of generated video. A 60-second clip costs $17–$30. RunwayML's Pro plan gives you ~90 seconds of Gen-4.5 for $28/month. mcp-video: $0, forever.
Your workflow depends on their API staying up, their pricing not changing, and their model versions not breaking your prompts. With mcp-video, FFmpeg runs on your machine — no API deprecation risk.
Every wasted turn costs money. With raw FFmpeg, Claude spends tokens parsing hundreds of lines of stderr, retrying failed commands, and re-probing files it already processed. mcp-video eliminates all of that.
| mcp-video | Claude Code + FFmpeg | |
|---|---|---|
| Turns for 5 operations | 4–5 | 6–10 |
| Output tokens per turn | 50–80 (JSON) | 200–500 (stderr) |
| Error retries | 0–1 (pre-validated) | 2–3 typical |
| Re-probing between steps | Never | Every step |
| Token reduction | ~40–50% fewer turns | baseline |
Every tool returns structured JSON. On success, you get output paths, duration, resolution. On failure, you get error types with auto-fix suggestions.
| Tool | What It Does |
|---|---|
| video_info | Get metadata: duration, resolution, codec, fps, file size |
| video_trim | Trim clip by start time + duration or end time |
| video_merge | Concatenate multiple clips with optional transitions |
| video_add_text | Overlay text, titles, captions with custom positioning |
| video_add_audio | Add, replace, or mix audio tracks with fade effects |
| video_resize | Change resolution or apply preset aspect ratios |
| video_convert | Convert between mp4, webm, gif, mov |
| video_speed | Speed up or slow down (slow-mo, time-lapse) |
| video_thumbnail | Extract a single frame at any timestamp |
| video_preview | Generate fast low-res preview for quick review |
| video_storyboard | Extract key frames + create grid for review |
| video_subtitles | Burn SRT/VTT subtitles into the video |
| video_watermark | Add image watermark with opacity control |
| video_export | Render final video with quality presets |
| video_edit | Full timeline-based edit (JSON DSL for complex edits) |
| video_extract_audio | Extract audio as mp3, wav, aac, ogg, or flac |
| video_crop | Crop to rectangular region with custom offset |
| video_rotate | Rotate 90/180/270 degrees or flip horizontal/vertical |
| video_fade | Add fade in/out effects to video |
| video_filter | Apply visual filter (blur, sharpen, grayscale, sepia, invert, vignette) |
| video_blur | Blur video with custom radius and strength |
| video_color_grade | Apply color preset (warm, cool, vintage, cinematic, noir) |
| video_normalize_audio | Normalize loudness to LUFS target (YouTube, broadcast, Spotify) |
| video_overlay | Picture-in-picture overlay with positioning and opacity |
| video_split_screen | Side-by-side or top/bottom layout for two videos |
| video_batch | Apply same operation to multiple files at once |
mcp-video works as an MCP server for AI agents, a Python library for scripts, and a CLI tool for the terminal.
Add to your Claude Code or Cursor MCP config. Then just talk to your agent.
{
"mcpServers": {
"mcp-video": {
"command": "uvx",
"args": ["mcp-video"]
}
}
}Then: *"Trim this video from 0:30 to 1:00, add a title, and export for TikTok."*
Clean API for automation, pipelines, and batch processing.
from mcp_video import Client editor = Client() clip = editor.trim("v.mp4", start="0:30", duration="15") final = editor.resize(clip.output_path, aspect_ratio="9:16") result = editor.export(final.output_path) print(result.resolution) # 1080x1920
# Get info $ mcp-video info video.mp4 # Trim and convert $ mcp-video trim video.mp4 -s 0:30 -d 15 $ mcp-video convert video.mp4 -f webm -q high # Generate storyboard $ mcp-video storyboard video.mp4 -n 12
Describe multi-track edits with video clips, audio, text overlays, transitions, and export settings — all in one call.
editor.edit({ "width": 1080, "height": 1920, "tracks": [ { "type": "video", "clips": [ {"source": "intro.mp4", "duration": 5}, {"source": "main.mp4", "trim_start": 10, "duration": 30}, ], "transitions": [{"after_clip": 0, "type": "fade", "duration": 1.0}], }, { "type": "audio", "clips": [{"source": "music.mp3", "volume": 0.7}], }, { "type": "text", "elements": [{"text": "EPISODE 42", "position": "top-center"}], }, ], "export": {"format": "mp4", "quality": "high"}, })
Pre-built templates handle the correct dimensions, text positioning, and export settings for each social media platform.
9:16, 1080x1920. Caption at bottom-center. Optional background music at volume 0.5.
9:16, title at top-center with size 42. Same format as TikTok but different text placement.
9:16, caption at bottom-center. Clean, simple template.
16:9, 1920x1080. Supports title card (3s), outro clip, and background music.
1:1, 1080x1080. Square format with optional caption overlay.
Build your own with the Timeline DSL. Any dimensions, any number of tracks.
Every model, error class, template, tool, and workflow is tested. Unit tests don't even need FFmpeg installed.
# 1. Install FFmpeg (if you don't have it) brew install ffmpeg # For full text overlay support (drawtext filter): # brew install freetype harfbuzz # brew reinstall --build-from-source ffmpeg # Verify: ffmpeg -filters | grep drawtext # 2. Install mcp-video pip install mcp-video # 3. Add to your Claude Code MCP config # (Settings > MCP Servers > Add > mcp-video) # 4. Start editing! # Just tell your agent what you want: # "Trim this video to 30 seconds and add a title"