Local video editing · Reversible decisions · AI through MCP
Live Showcase & Downloads · Get started · Connect your AI · Read the docs · Build with us
Reeltern gives people and AI assistants the same editing engine. Select moments from a supplied transcript, arrange source clips, review every decision, and export a video with a verification receipt. Originals stay intact. Earlier edits stay available.
The idea: an AI edit should come with source evidence, a reversible plan, and a result you can check.
Note
Early preview, actively being built. The core suite passes all 59 automated tests locally (including live in-window PreviewEngine compositing, blur_fill portrait pixel halos, animated progress bar & kinetic ASS pixels, Ken Burns camera motion, sidechain-ducked music mix, desktop Studio NLE workstation actions, and stdio MCP checks). Optional enhancement adapters remain experimental. See exactly what is tested →
| Create & shape | Stay in control | Finish with evidence |
|---|---|---|
Direct & craft at high speed. Auto-compose with compose_video / reeltern compose (cinematic, viral_reel, tutorial, documentary, noir_trailer), title cards, blur_fill vertical halo framing, and background music. |
Keep every revision. Split at the playhead, animate (zoom_in/zoom_out/pan_left/pan_right), grade color (10 looks incl. teal_orange, studio, vintage), mix audio (voice_cleanup + sidechain ducking), and undo cleanly. |
Self-critique & verify. Inspect PNG contact sheets (reeltern://visual/...), run closed-loop audit_render (EBU R128 loudness, exposure, black frames), and check SHA-256 receipts. |
Work with your assistant. Use 31 AgentFlight Grade-A MCP tools (0 warnings), PNG visual inspection resources, or a compacted local Ollama model. |
Carry captions forward. Retime subtitles to cuts/speed and deliver soft tracks, kinetic word-highlight burned-in ASS styles, or both. |
Reuse completed work. Cache unchanged segments, contact sheets, and waveforms cleanly and retry interrupted jobs. |
Desktop, command line and MCP all use the same core. No Reeltern account or API key is required. Your chosen assistant has its own access requirements and data policies.
flowchart LR
A["Footage + supplied text"] --> B["Human or AI draft"]
B --> C["Review source ranges"]
C --> D["Captioned preview"]
D --> E["Verified export"]
C -->|"Save a revision"| B
Every plan records what was selected, where it came from, how fast it plays, and why it was chosen. Model explanations remain separate from actual source quotes. A valid plan still needs editorial judgment.
Prefer a download? GitHub Releases provides standalone Desktop Studio GUI apps for macOS (Apple Silicon & Intel), Windows, and Linux — plus Python installation bundles, wheel, and source archives. Choose a package →
Tip
First launch on macOS / Windows (unsigned open-source builds): If macOS says "Apple could not verify Reeltern Studio", copy Reeltern Studio.app to Applications, click Done, then open System Settings → Privacy & Security → Open Anyway. On Windows SmartScreen, click More info → Run anyway — or run reeltern studio directly via Python below.
Try a complete edit before bringing your own footage. Use Python 3.12 for parity with the test environment; the package supports Python 3.10+.
git clone https://github.com/codesphere-u/reeltern.git
cd reeltern
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[mcp]'
# Generate footage, make a captioned edit, and verify the result.
reeltern demo --workspace demo-projectWindows / PowerShell
git clone https://github.com/codesphere-u/reeltern.git
cd reeltern
py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[mcp]"
.\.venv\Scripts\python.exe -m reeltern.cli demo --workspace demo-projectUse .\.venv\Scripts\python.exe -m reeltern.cli in place of reeltern for the commands below. Activation is optional.
What the demo produces: six seconds of synthetic source footage become a four-second preview. The middle section is omitted, captions are retimed, and a second render demonstrates segment reuse. Open the video path printed by the command. No account, user footage, or model download is needed.
FFmpeg is supplied by imageio-ffmpeg on supported platforms; an installed system FFmpeg is preferred. Desktop mode additionally needs Tkinter. Installation help →
reeltern studioImport or Load Demo → AI Director or Trim → Review Timeline → Audit QA → Export.
Use Load Demo to generate sample footage and inspect a multi-clip timeline immediately, Connect AI to copy Antigravity / Claude Code / Codex / Cursor / VS Code MCP settings, AI Director for one-click composition, or Transcript & AI to import an SRT. Playback opens in your default video player.
# Auto-compose a vertical social edit with music ducking, kinetic polish, and QA audit.
reeltern compose footage.mp4 --style viral_reel --music bgm.wav --preview --audit
# Or inspect and compose manually around audible activity.
reeltern edit footage.mp4 --workspace my-project --preset portrait --fit blur_fill --previewQuiet audio is a signal, not a judgment about meaning. Review the proposed cuts before relying on them for your final edit.
reeltern transcript interview.mp4 interview.srt --workspace my-project
# Optional: ask an already-installed local Ollama model for a draft.
reeltern agent TRANSCRIPT_ID "Keep the explanation and conclusion, preserving qualifications" \
--workspace my-project --model YOUR_INSTALLED_MODEL --seconds 45Use the transcript ID printed by the first command. An MCP assistant can select transcript cues directly without Ollama. Reeltern does not yet transcribe speech. Transcript workflow →
Antigravity · Claude Code · Codex · Cursor · VS Code · other local MCP clients
Generate configuration from the Python environment where Reeltern is installed (or click Connect AI inside Reeltern Studio):
reeltern mcp-config antigravity \
--workspace editing-project \
--media-root demo-project/demo-inputReplace antigravity with claude-code, codex, cursor, vscode or generic. The command prints settings with absolute paths; merge the Reeltern entry into your client's configuration. It does not change settings automatically. Use an existing footage folder as --media-root.
Then ask your connected assistant:
Import synthetic-source.mp4. Keep source seconds 0–2 and 4–6, in that order. Create a portrait preview, preserve the whole image, wait for completion, and show me the verified result.
| Connection | Editing workflow | Tool contracts |
|---|---|---|
| Client setup & configuration | Agent operating guide | 31-tool reference |
| Ready-to-adapt templates | Executable MCP example | Machine-readable schemas |
The server uses local stdio, with imports restricted to the configured media root. Give simultaneous clients separate workspaces. Cloud assistants may receive transcript text, prompts and tool results under their own policies; local rendering does not make every assistant interaction private.
Explore the engineering behind an edit
| Mechanism | What it provides |
|---|---|
| Content-addressed source snapshots | Imported media is copied; source hashes are checked during analysis and rendering. |
| Immutable cut plans | Source ranges, clip order, speed and output timing are saved with revision links. |
| Bounded local agent | Selects existing transcript cue IDs, checks the duration budget, and allows at most one repair attempt. |
| Typed MCP inputs | Assistants discover tool contracts; nested clip/cue schemas reject unknown fields. |
| Segment cache | Unchanged encoded segments are reused after checksum validation. |
| Persistent render jobs | Progress, cancellation, errors, interrupted state and retry remain inspectable. |
| Verification receipts | Completed output passes dimension/duration checks, full video/audio decoding and source-integrity checks. |
One engine serves all interfaces. No arbitrary shell command, HTML execution or public network endpoint is exposed through MCP. Technical verification does not establish aesthetic quality or perfect perceptual sync.
Reeltern brings a shared workflow around established open-source media tools. FFmpeg powers the tested core. The optional toolkit preserves experimental adapters for the wider ecosystem:
| Workflow | Upstream projects |
|---|---|
| Find and shape moments | PySceneDetect · Auto-Editor · librosa |
| Refine the image | vid.stab · color-matcher · rembg |
| Interpolate and enhance | RIFE · Real-ESRGAN |
| Separate sound and create graphics | Demucs · HyperFrames |
| Process and export | FFmpeg |
Launch it with reeltern toolkit. These adapters need separate dependency/model setup and acceptance testing; they are not exposed through MCP. You can import a completed adapter output into the editor.
Upstream credits & source review · Adapter setup & limitations
Help make the first useful edit easier—and the next one better. There is room for media engineers, Python developers, AI evaluators, designers and people who edit video every day.
| If you care about… | A useful contribution |
|---|---|
| Editing by meaning | Optional transcription, word alignment, and evaluations that catch lost context. |
| A better editing experience | In-app preview, waveform selection, keyboard access and clearer review flows. |
| Reliable exports | Real-footage tests for phone rotation, variable frame rates, audio offsets and subtitles. |
| A smoother first run | Installers, dependency diagnostics, reproducible examples and clearer documentation. |
| Advanced media workflows | Promote an optional adapter through a tested, cancellable, documented integration. |
These are contribution opportunities, not released features. The roadmap gives priorities and acceptance criteria.
python -m unittest discover -s tests -v
python scripts/check_docs.pyThe suite generates its own media and exercises the real encoder, decoder and MCP transport. Documentation checks validate local links and configuration examples; protocol tests compare the published tool schemas with the running server.
Contributor guide · Implementation guide · Report a bug · Propose a feature
Coding with an assistant? Start with AGENTS.md, CLAUDE.md and the compact documentation index.
Important
This is an evolving local editor designed for reproducible automation and hands-on timelines. There is no built-in speech recognition, face tracking, one-click installer, or hosted service.
Formats, effects, captions, resource limits and recovery
- Create & polish: Generate title/chapter cards (
create_card), auto-compose production timelines with AI Director (compose_video/compose), split clips non-destructively (split_clip), inspect PNG contact sheets, scene cuts & audio energy peaks, and apply 10 color grades (cinematic,warm,crisp,noir,vivid,moody,teal_orange,vintage,studio), Ken Burns camera motion (zoom_in,zoom_out,pan_left,pan_right), fades,0.25×–4×speed, rotation/flip, lower-third overlays, sliding bottom progress bars,voice_cleanupEQ/compressor,sidechaincompressducked music beds (import_audio), EBU R128 loudness normalization, and 10 transitions (crossfade,wipeleft/right,slideleft/right,dissolve,smoothleft,circleopen,fadeblack). - Output: SDR, 30 fps, H.264/AAC MP4. Landscape
1280×720, portrait720×1280, square720×720,fhd1920×1080,reels1080×1920; previews halve both dimensions.containpads;covercenter-crops;blur_fillcenters sharp foreground over a Gaussian-blurred background canvas. - Streams: only the first source video/audio streams are used, plus an optional attached background audio track. Original HDR, metadata, subtitle tracks, chapters and multitrack audio are not preserved.
- Captions: supplied transcripts produce SRT sidecars and support selectable MP4 subtitles (
soft), stylized ASS burned-in pixels (burninclean/boxed/bold/kinetic), orboth.kineticgroups existing cue words into timed phrase/word highlights; mid-cue trims still retain the full cue text when word alignment is absent. - Speed: 0.25×–4× uses frame-rate conversion and chained FFmpeg
atempofilters; it does not invoke RIFE interpolation. - MCP token efficiency: Schemas are compacted cleanly (
100/100 Grade Aon AgentFlight linter,0warnings and0errors across 31 tools);EditorialAgentsends compact cues and relies on Ollama structuredformatoutput. - Resources: input limit 8 GiB/file, timeline limit 200 clips/four hours, queue limit eight unfinished renders. Snapshots and caches consume disk space; automatic eviction is not implemented.
- Recovery: the next renderer marks unfinished jobs interrupted. Explicit retry creates a new job and can reuse verified segments; there is no automatic restart or mid-encoder resume.
- Cancellation: render jobs are cancellable; import/analysis cancellation is not implemented. A client timeout does not prove background work stopped.
- Trust: workspaces are trusted local state. Models can still make poor editorial choices. Core tests use synthetic fixtures; broad real-footage and every-client UI acceptance remain open work.
Your footage. Your decisions. A workflow you can build on.
GPL-3.0-or-later · Built on open-source media software · No bundled model weights
Third-party engines and model weights retain their own licenses. See upstream credits.