AI-native · Windows · Open source

The AI-operatedvideo editor for Windows.

You describe the edit. It makes it. Kaestral watches your footage, hears every word, and cuts.

Download for Windows View on GitHub

Free·GPLv3·Works with the Claude Code you already have

kaestral — editing from a prompt
What it does

Three things, done for you.

01 — DIRECT

AI-operated. You direct, it edits.

Type “cut the boring parts, add captions, punch in on the hook” — and watch your timeline change. Every action runs through the editor’s own tools, so you get a real, editable project, not a black-box render. Drive it with Claude Code (free) or the in-app chat.

Kaestral editor: multi-track timeline with media, preview, and transport controls
02 — CONDENSE

Raw footage → publish-ready reel.

Drop a 20-minute recording; get a tight, captioned, beat-cut short. Kaestral removes filler words, finds the hook, and captions on the word — automatically. Transcription runs on-device; cuts land on silences and beats.

03 — CREATE

Make SaaS & product videos without an editor.

Animated intros, logo reveals, data-viz, and transition stingers — generated from a sentence and rendered onto your timeline. Launch videos, demos, and ads, without hiring an editor.

Real pro timeline — multi-track, keyframes, 16 blend modes Vision — the AI sees your frames Transcription — word-level, on-device Beat / silence detection Palette extraction Color — wheels, curves, LUTs Motion graphics — Remotion Import from a URL Export — H.264 / H.265 / ProRes + Premiere / Resolve 51 MCP tools
Getting started

Up and running in three steps.

No project setup, no timeline tutorial. Install, connect, and start describing the edit you want.

1

Install Kaestral

Download the Windows installer and run it. The editor and its engine start together — nothing else to configure.

2

Connect Claude

One command connects the Claude Code you already have. Requires Node/npm and FFmpeg on your PATH.

claude mcp add kaestral -- npx kaestral
3

Describe your edit

Type what you want in plain language — “cut the dead air and add captions” — and watch the timeline do it.

Jobs to be done

What people make with it.

Real, shipped capabilities across eight kinds of work — grouped by who's using them. Everything here runs today, except the last group.

🎬

Short-form / Creator

  • Turn a 20-minute recording into a tight 60-second reel — auto-removes filler words, dead air, and false starts
  • Cut a podcast into viral clips — finds the hook, trims to the payoff
  • Word-synced karaoke captions — every word highlighted on the syllable, not near it
  • Cut on the beat — automatically syncs cuts and animations to the music's BPM
  • Reframe landscape → 9:16 vertical for Reels, Shorts, TikTok
  • Punch-in zooms on key moments
  • Generate a viral-style reel from raw footage with one prompt
  • Platform-specific delivery — Reels / Shorts / TikTok / YouTube / feed
📊

SaaS / Product / Startup

  • Make a product demo without an editor — browser chrome, arrows, highlight boxes, spotlight-dim
  • Launch films — cinematic intros with logo reveals and specular sweeps
  • Animated data-stories — bar/line/area charts that grow, numbers that count up
  • Feature walkthroughs — annotated screen recordings
  • Sales & pitch videos
  • Investor update videos
  • Changelog / release videos
  • Onboarding & tutorial videos
  • Ad creative — hook → beats → CTA structure
✂️

Editing (the classic NLE)

  • Multi-track timeline (video, audio, image, text, Lottie)
  • Trim, split, ripple delete, drag, snap, razor
  • Speed / retime, keyframe animation with easing lanes
  • Transforms, crop, opacity, blend modes
  • Transitions — cross-dissolve, wipe, push, glitch, RGB-split
  • Layouts — picture-in-picture, split-screen, grid
  • Color grading — exposure, contrast, saturation, temperature, palette-matched grades
  • Audio — waveforms, volume, fades, mixing
  • Export — H.264 / H.265 / ProRes, 1080p or 4K
🧠

AI-Operated (the differentiator)

  • “Cut the boring parts” — silence and pause detection
  • “Remove every ‘um’” — word-level surgical cutting
  • “Caption this” — word-accurate, styled, animated
  • “Make this a viral reel” — reframe → hook → cut → caption → grade
  • “Pull my brand colors from this clip” — palette extraction → auto-grade
  • “Animate an intro that says [X]” — bespoke motion graphics from a sentence
  • “Find the best moment” — it watches your frames and picks
  • Claude designs the whole film — art-directed compositions, not templates
🔌

Developer / MCP

  • claude mcp add kaestral -- npx kaestral — one-line install
  • 52 tools exposed to any LLM (Claude Code, Cursor, Codex, ChatGPT)
  • Drive video editing from your terminal — scriptable, agentic
  • Automate video pipelines — batch, CI, repeatable renders
  • Build on it — open source, GPLv3
📥

Import / Interop

  • Import from URL (paste a link, it pulls the video)
  • Drag-and-drop any video / image / audio
  • Open .kaestral projects
  • Export XMEML → Premiere Pro
  • Export FCPXML → DaVinci Resolve / Final Cut
  • SRT caption export
🔒

Privacy

  • Everything stays local. Your video never leaves your machine.
  • Local transcription (whisper), local rendering, local projects
  • No cloud upload, no account required
🔮

Pro

Coming — join the waitlist
  • AI video & image generation — prompt → new footage on your timeline
  • Translation & dubbing — Hindi + regional Indian languages, and beyond
  • AI upscaling — temporally consistent, no flicker
  • Advanced generative motion — PixiJS particles, shaders
Full capability surface

51 MCP tools. Any LLM can drive them.

Claude Code, Cursor, Codex, ChatGPT — any client that speaks MCP can call every tool below. This is the complete surface, not a marketing sample.

👁️

Perception — it sees, hears, and feels rhythm

The tools that give the AI eyes and ears before it ever touches your timeline.

ToolWhat it does
see_videoExtracts and returns real frames — the LLM watches your footage
get_transcriptWord-level transcription (local whisper), exact timestamps
analyze_audioBeats, BPM, onsets, silence and pause detection
extract_paletteDominant colors → brand-matched grading
inspect_mediaReal ffprobe metadata — dimensions, fps, duration, audio
inspect_timelineStructured read of the live timeline state
search_mediaSemantic search across your library

Traditional NLEs have none of this. This is why Kaestral edits with judgment instead of guessing.

✂️

Timeline editing

get_timelineadd_clipsinsert_clipsremove_clipsmove_clipssplit_clipsripple_delete_rangesset_clip_propertiesset_keyframesremove_trackssync_audioundo
  • Multi-track (video/audio/image/text/Lottie)
  • Trim, split, ripple, razor, drag, snap
  • Speed/retime, transforms, crop, opacity, blend modes
  • Keyframe animation with easing lanes
  • Linked A/V, sync-lock, full undo history
🗣️

Word-level editing (the killer feature)

ToolWhat it does
remove_wordsSurgically cut specific words — “remove every ‘um’”
add_captionsWord-synced karaoke captions, styled and animated
add_texts / update_textFull typography: font, size, weight, color, outline, background, alignment, animation presets
🎨

Motion graphics — traditional NLEs have zero of this

ToolWhat it does
compose_motionThe art-direction surface. LLM authors bespoke films — beats, layers, per-property animation, bezier easing, hold-and-settle, optical placement, camera moves, transitions
generate_titleFast animated title cards (canvas engine)

Primitives available to the LLM:

text (spring/typewriter/word-reveal/kinetic/karaoke)shapeshairlineswaveformsanimated timelineslogo marksbar/line/area chartscountersparticlesglow fieldsgridsmasks & revealssplit/grid layoutsdepth-of-fieldmotion blurtext-on-pathlighting sweepsKen Burnscountdown & glitch stingersScreenMock (browser chrome)callouts (arrow, highlight box, pointer, spotlight-dim)Video/Image compositing
  • Camera: push-in · pan · rack-focus · parallax
  • Transitions: wipe · dissolve · push · glitch · RGB-split · cut
  • Brand tokens: automatic on-palette output

Traditional NLEs have none of this. Motion graphics from a sentence, art-directed — not a template picker.

🎛️

Color & effects

apply_colorapply_effectinspect_color
  • Exposure · contrast · saturation · temperature · tint
  • Lift / gamma / gain · curves · LUTs
  • Blur · sharpen · stylize · blend modes
📐

Layout

apply_layout

Picture-in-picture · split-screen · grid · side-by-side, auto-composed.

📁

Media library

import_mediaimport_from_urllist_folderscreate_foldermove_to_folderrename_mediarename_folderdelete_mediadelete_folder

import_from_url — paste a link → it pulls the video.

📤

Export

export_project
  • video → H.264 / H.265 / ProRes, 1080p or 4K
  • xml → XMEML → Premiere Pro
  • fcpxml → DaVinci Resolve / Final Cut
  • kaestral → project package

Plus: SRT caption export.

🧠

Skills — the craft layer

list_skillsread_skill

Bundled playbooks the LLM auto-loads:

  • art-direction — the master motion-designer playbook: decision process, optical composition, rhythm, restraint, trade-offs, worked examples, failure modes
  • viral-reel — hook → cut → punch-in → caption → grade
  • beat-sync-cutting — cut to the music
  • creative-director — palette, brand, typography
  • caption-styles — word-pop, karaoke, bold-boxed, typewriter
  • broll-planner — B-roll placement
  • promo-ad — hook → beats → CTA
  • platform-delivery — Reels / TikTok / Shorts / YouTube specs
⚙️

Project

set_project_settingsexport_projectlist_modelssend_feedback

Local .kaestral projects.

🔮

Pro

Stubbed — waitlist open
generate_videogenerate_imagegenerate_audioupscale_media

Translation & dubbing (Hindi + regional Indian languages) · temporally-consistent upscaling · PixiJS shaders.

🔒

Everything runs local

Local whisper · local rendering · local projects · your video never leaves your machine.

Why Kaestral

An AI editor with eyes and ears.

Native to Windows, free, and open source. It doesn’t just take commands — it watches your frames, hears every word, and edits through a real, editable timeline.

 KaestralWhat that means
PlatformWindowsNative desktop app · macOS on the roadmap
AI editing over MCP✓Drive it from the Claude Code you already have
Perception✓ on-deviceTranscribe · see frames · detect beats & silence — locally
Real timeline✓Multi-track, keyframes, blend modes — a true editable project
PriceFree · GPLv3Open source; yours to keep and modify
AI generation⏳ Pro (waitlist)Prompt-to-clip is coming to the Pro tier

Kaestral doesn’t compete on generation today — its edge is Windows + real perception + a free, open, Claude-Code-native workflow. Generation lands later, in the Pro tier.

Privacy

Your footage never leaves your device.

Kaestral is local-first. Your video, audio, and project files are processed and stored on your machine — never uploaded to a server. Transcription, vision, and beat detection all run on-device.

On your disk

Footage & projects — read, edited, and rendered locally. Nothing is synced to the cloud.

On your CPU/GPU

Transcription, vision, beat & silence detection — all run on-device, not through a remote API.

Open source

GPLv3 — the code doing this is auditable. See exactly what runs and where.

Two things — and only two — leave your device: your AI prompts, sent to Anthropic’s API if you use in-app chat with your own key; and your email, sent to our form provider only if you join the Pro waitlist below. Everything else — your video, audio, and project files — stays on your machine.

For developers

MCP-first. One line in.

Kaestral is an MCP server. Point the Claude Code you already have at it and start editing by prompt — no new app to learn.

# one command — connects Claude Code over stdio
claude mcp add kaestral -- npx kaestral
claude

# then just ask
# “cut the silent parts of demo.mp4,
#  add captions, and export it.”

# prefer HTTP? the editor app also exposes:
# claude mcp add --transport http kaestral \
#   http://127.0.0.1:19789/mcp

Claude Code

claude mcp add kaestral -- npx kaestral — one command, no server to run manually.

Cursor

Add to ~/.cursor/mcp.json: "kaestral": { "command": "npx", "args": ["kaestral"] } under mcpServers.

Claude Desktop

Settings → Developer → Edit Config (%APPDATA%\Claude\claude_desktop_config.json), same npx kaestral entry — or double-click the one-click .mcpb from Releases.

In-app chat

No MCP client needed — bring your own Anthropic key and edit from the app window.

ChatGPT?

ChatGPT connectors need a hosted/remote server; Kaestral is local-first. Use Claude Code, Cursor, or Claude Desktop — cloud support is on the roadmap.

Compatibility

Standard MCP over stdio (or HTTP) with a stable ~51-tool contract. Requires Node/npm + FFmpeg on PATH. Exports Premiere (XMEML) and Resolve / FCP (FCPXML).

Kaestral Pro

AI generation is coming.

Type a prompt, get a real clip on your timeline. Generate video, images, and B-roll inside Kaestral. Join the waitlist — we’ll email you the moment it opens.