SkyCut
SkyCut is an MCP server for turning raw drone footage into polished, AI-assisted marketing cuts through a conversational interface.
Check dependency health (ffmpeg, encoders, API key)
List mounted volumes to pick a source drive
Initialize a read-only project from a footage folder
Scan footage: index metadata and build 720p proxies
Analyze footage with vision to build a searchable footage graph (subject, movement, quality, aesthetic scores)
Search moments by subject, movement, stability, exposure, aesthetic score, or text
Get timeline versions and shot lists
Propose an AI-directed cut from a creative brief and target duration
Apply structured timeline edits (insert, remove, reorder, retrim, transitions, music) as immutable versions
Render fast 720p previews from proxies (works without source drive)
Render full-quality HEVC finals from original files (requires explicit version and mounted drive)
Check project status: pipeline state, versions, renders, drive mount
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@SkyCut scan the drone footage on my USB drive"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
SkyCut
Local-first MCP server (stdio, TypeScript) that turns a folder of raw drone footage into a polished marketing cut, driven conversationally from Claude Code or Claude Desktop:
scan → analyze (vision) → footage graph → propose cut → human approve → render
Point it at a USB drive of raw drone video and produce a 60–90 second marketing edit with crossfades and an optional music bed. The source drive is treated as read-only; all derived artifacts (proxies, frames, timelines, renders) live in ~/SkyCut/projects/<slug>/ on the internal disk.
Requirements
Node 20+
ffmpeg + ffprobe —
brew install ffmpeg(Apple Silicon videotoolbox hardware encoders used for all renders)ANTHROPIC_API_KEY— for vision analysis and cut proposal
Related MCP server: MCP Research Tools
Install
npm install
npm run buildClaude Code
claude mcp add skycut --env ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY -- node /absolute/path/to/skycut/dist/index.jsClaude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"skycut": {
"command": "node",
"args": ["/absolute/path/to/skycut/dist/index.js"],
"env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
}
}
}Verify with MCP Inspector
npm run inspectorWorkflow
A typical session, in plain conversation with Claude:
skycut_list_volumes()— find the USB driveskycut_init_project("/Volumes/DJI/lodge-trip")— create the workspaceskycut_scan_footage()— index clips, build 720p proxies (idempotent; unplug the drive afterwards if you like)skycut_analyze_footage()— Claude vision scores every 4s keyframe into a searchable footage graph (cached; asks before runs over 500 frames)skycut_propose_cut("Cinematic 90-second marketing cut, slow build, reveal the lodge at golden hour", 90)— AI director assembles timeline v1skycut_render_preview()— fast 720p render from proxies; watch itskycut_apply_timeline_edit({ edits: [...] })— revise; every edit is a new immutable versionskycut_render_final(2)— full-quality HEVC render from the originals (drive must be mounted)
Explore in between with skycut_search_moments({ subject: "lodge", min_aesthetic: 7 }) and skycut_project_status().
Web UI
A local chat UI (no auth — local use only) wraps the same core pipeline in a browser:
ANTHROPIC_API_KEY=sk-ant-... npm run web # → https://dev.ecoworks.ca:3080Chat agent — a Claude-driven agent with tool use over the SkyCut core (in-process, no MCP hop): propose cuts, search moments, apply edits, and render, with SSE-streamed progress bars and per-turn/session cost tracking
Footage panel (📁) — browse volumes/folders and open one as a read-only project source, drag-and-drop video uploads (into
~/SkyCut/uploads), scan + analyze with live progress and the vision cost gate, and switch between projects (chat archives on switch)Timeline panel (🎞️) — visual shot strip with keyframe thumbnails; drag shots to reorder, drag shot edges to retrim (every drop saves a new immutable version); click a shot for trim/speed/source details; compare any two versions with a color-coded diff; a playhead tracks the matching preview player and clicking a shot seeks it
Music (🎵) — search royalty-free tracks (Jamendo), preview in-chat, download into
~/SkyCut/musicRenders — previews and finals embed as inline players (same-origin media with HTTP range support)
About — in-app changelog, how-it-works guide, and roadmap
The web server operates on the active project (the one last opened with skycut_init_project). Chat history persists across restarts; edits made in the panel and in chat share one code path and one version history.
Tools
Tool | Purpose |
| Dependency status (ffmpeg, encoders, API key) |
| Mounted volumes with free space |
| Create/reopen a project; source is read-only |
| ffprobe metadata + manifest + SQLite + 720p proxies |
| Vision analysis → scored segments (footage graph) |
| Query segments by subject/movement/quality/score/text |
| AI director → validated timeline version |
| Timeline JSON + shot list |
| Structured diff → new immutable version |
| 720p from proxies |
| Up-to-4K HEVC from originals; explicit version required |
| Pipeline state + drive mount check |
Guardrails
The USB source is never written to; unplugging it mid-session degrades gracefully (proxies keep previews working)
Timeline versions are immutable — edits always create
v<N+1>Nothing is ever auto-finalized:
skycut_render_finalrequires an explicit versionEvery ffmpeg command is logged to the project's
logs/ffmpeg.logVision runs needing more than 500 frames report an estimated cost and wait for confirmation
Development
npm run build # tsc (the web server imports from dist/ — rebuild before `npm run web`)
npm test # vitest — vision/director mocked; ffmpeg integration uses generated testsrc clips
npm run inspector # poke the MCP server interactively
npm run web # chat UI on https://dev.ecoworks.ca:3080 (needs ANTHROPIC_API_KEY)Documentation
docs/README.md — architecture, design decisions, workspace layout, troubleshooting, extending
docs/TOOLS.md — complete tool API reference, timeline schema, edit operations, error conventions
CHANGELOG.md — version history
SkyCut Overview deck — 10-slide presentation: the problem, pipeline, footage graph, director, guardrails, both front ends, stack and roadmap (private Claude artifact — ask the owner for access)
prompt_plan.md(original spec) andprogress.md(per-phase build history)
Available Tools
12 toolsskycut_analyze_footageAnalyze footage (vision)AIdempotent
Sample keyframes (1 per 4s) from every scanned clip and analyze them with Claude vision to build the footage graph: scored segments with subjects, camera movement, and quality flags. Cached — already-analyzed clips are skipped unless force=true. If the run needs >500 frames, returns a cost estimate instead; re-call with confirm=true to proceed.
Inputs: force, confirm.
skycut_apply_timeline_editEdit timelineA
Apply structured edits to the latest (or given) timeline version, or replace it wholesale. Versions are immutable — the result is always saved as a NEW version. Ops: insert {at_index, clip}, remove {id}, reorder {id, to_index}, retrim {id, in_s?, out_s?, speed?}, set_transition {id, transition|null}, set_music {music|null}. Example: { edits: [{ op: "retrim", id: "c3", out_s: 51.0 }] }
Inputs: edits, timeline, base_version.
skycut_get_timelineGet timelineARead-only
Return a timeline version (latest by default) as JSON plus a human-readable shot list.
Inputs: version.
skycut_healthSkyCut health checkARead-only
Report status of required dependencies (Node, ffmpeg, ffprobe, videotoolbox encoders, ANTHROPIC_API_KEY). Run this first if any other tool returns unexpected errors.
Inputs: no named parameters.
skycut_init_projectInitialize projectAIdempotent
Create (or reopen) a SkyCut project for a folder of source footage and make it the active project. The source path is treated as READ-ONLY; all derived files go to the workspace under ~/SkyCut/projects/. Example: skycut_init_project({ source_path: "/Volumes/DJI/ogoki-lodge" })
Inputs: source_path (required), name.
skycut_list_volumesList mounted volumesARead-only
List volumes under /Volumes with free/total space (GB), to help pick the source drive for skycut_init_project.
Inputs: no named parameters.
skycut_project_statusProject statusARead-only
Pipeline state for the active project: scan/analysis progress, timeline versions, renders, and whether the source drive is currently mounted.
Inputs: no named parameters.
skycut_propose_cutPropose cutA
Have the AI director assemble a timeline from the footage graph: narrative arc, varied movement, 3-8s shots, crossfades, within ±5% of the target duration. Saves an immutable timeline version for review — NEVER renders anything itself. Example: skycut_propose_cut({ brief: "90s cinematic cut for a fly-in fishing lodge, slow build to reveal", duration_s: 90 })
Inputs: brief (required), duration_s (required), style, music_path.
skycut_render_finalRender finalA
Full-quality render (HEVC videotoolbox, up to 4K, AAC audio) from the ORIGINAL files on the source drive. Requires an explicit timeline version — never renders 'latest' implicitly. The drive must be mounted.
Inputs: version (required).
skycut_render_previewRender previewA
Fast 720p preview render of a timeline version (latest by default) from proxies — works even with the source drive unplugged. Output goes to the project renders/ folder.
Inputs: version.
skycut_scan_footageScan footageAIdempotent
Recursively index video files (.mp4/.mov/.mts/.mkv) under the active project's source path: ffprobe metadata into the footage database + manifest.json, and build 720p proxies for previews. Idempotent — re-running skips clips and proxies that already exist. The source is never modified.
Inputs: no named parameters.
skycut_search_momentsSearch momentsARead-only
Query the footage graph for segments by subject, camera movement, quality, aesthetic score, or free text against notes. Example: { subject: "lodge", movement: "orbit", min_aesthetic: 7 }
Inputs: text, limit, subject, exposure, +3 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 12 tools
Each tool targets a distinct step in the video editing pipeline (init, scan, analyze, search, propose, edit, render, health), with clear descriptions preventing overlap. Even similar tools like scan and analyze are clearly distinguished by their operations and outputs.
All tools use the 'skycut_' prefix and mostly follow verb_noun pattern. Exceptions like 'skycut_health' and 'skycut_project_status' use nouns, but the overall pattern is consistent enough for easy navigation.
12 tools are well-scoped for a video editing MCP server, covering the full pipeline from setup to final render without unnecessary redundancy. Each tool serves a clear purpose.
The tool set covers the major workflow stages (init, scan, analyze, search, propose, edit, preview, render). Minor gaps exist (e.g., no standalone clip trimming or subtitle export), but the core pipeline is complete.
Related MCP Connectors
MCP server for OpenAI Sora AI video generation
Hosted YouTube MCP server for Claude, ChatGPT and Cursor, with each action switched per account.
MCP server giving Claude AI access to 22+ NYC public-record databases for real estate due diligence
Make animated videos by asking Claude, Cursor or any MCP client. Your AI writes it, we render it.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceA local MCP server that gives Claude Desktop full video editing capabilities via FFmpeg, Whisper, and yt-dlp.-
- AlicenseNot gradedqualityDmaintenanceLocal-first MCP server that gives Claude Code web search, page reading, video transcription, and image analysis — without paid API keys. Runs SearXNG + whisper.cpp natively on Apple Silicon for zero-cost, low-latency research workflows.1MIT
- FlicenseAqualityDmaintenanceA local-first MCP server that enables AI coding assistants to map codebases, transcribe onboarding videos with keyframe screenshots, and maintain structured Obsidian knowledge vaults directly from IDE chat.5-
- AlicenseNot gradedqualityDmaintenanceAn MCP server that gives Claude a conversational video-editing toolkit — trim, crop, color grade, caption, beat-sync, stabilize, and more — built on ffmpeg, faster-whisper, and yt-dlp, with optional DaVinci Resolve and CapCut integration.2MIT