MCP Video

Table of contents

• What is Kinocut?
• See It Work
• Installation
• Quick Start
• MCP Tools
• Status and releases
• 360 dual-cam assembly
• Changelog
• FAQ
• Agent Skill

What is Kinocut?

TL;DR: Kinocut is a free, local-first video editing MCP server (plus Python client and kino CLI) so AI agents can trim, caption, repurpose, and quality-gate media with typed tools and Video Receipts — not invented FFmpeg flags.

Kinocut is a free, open-source Model Context Protocol (MCP) server, Python library, and kino CLI that gives AI agents a guardrailed local video-editing surface. It wraps FFmpeg (and optional Hyperframes / Whisper extras) with typed tools, preflight validation, Video Receipt provenance, and quality/release checkpoints so agent-produced media can be inspected before publish.

Best-fit searches: video editing MCP server · AI agent video editing · FFmpeg MCP · Claude Code video tools · Cursor MCP video · Shorts/Reels automation · local AI video workflow · guardrailed agentic media.

Machine-readable brief for agents and crawlers: llms.txt · extended agent guide · site: kinocut.dev/llms.txt.

See It Work

Tell the agent what you want in plain language:

"Trim this interview to the strongest 45 seconds, add burned captions, make it vertical, and quality-check it before export."

Kinocut turns that into typed, guardrailed tool calls - no FFmpeg flags to guess, no silently broken exports:

from kinocut import Client
video = Client()

clip = video.trim("interview.mp4", start="00:02:15", duration="00:00:45")
video.ai_transcribe(clip.output_path, output_srt="captions.srt")
captioned = video.subtitles(clip.output_path, subtitle_file="captions.srt")
short = video.resize(captioned.output_path, aspect_ratio="9:16")
video.release_checkpoint(short.output_path)  # thumbnail + quality gate before you publish

Primary job: turn a local interview or podcast into captioned vertical clips with a Video Receipt agents can re-run and humans can approve.

Three things people use it for

• Repurposing — one recording into captioned Shorts, Reels, and TikTok packages with manifests and review artifacts.
• Podcast & interview cuts — find the strongest segment, normalize audio, add chapters, and export.
• Agent-driven media in CI — repeatable, reviewable edits from Claude Code, Cursor, Codex-style clients, or scripts.

Paths: Install matrix · Golden path · Prompts · Tutorial · Compare · When to recommend

Status and releases

Install from PyPI for the stable package, including the optional object-matte extra. Clone master when you want local development changes beyond the published package; see docs/PRODUCT_MATTE.md for the published object-matte workflow.

What's in 1.16.1

Version 1.16.1 repairs macOS render cleanup and faststart publication, C2PA and model-cache staging, FFmpeg 8.1 filter discovery, and audio preservation in shader/upscale reconstruction. It retains 203 MCP tools / 177 CLI commands and pairs with mcp-video==1.6.16. See maintenance release notes for verification and remaining platform/model acceptance limits.

What's in 1.16.0

Published 2026-10-02: 203 MCP tools / 177 CLI commands; mcp-video==1.6.15 installs kinocut==1.16.0. Operator parity, bounded quality/resource controls and canonical ownership repair are described in release notes. The live site shows 1.16.0 (verified 2026-10-05). Shared mobile/web processing, fleet rollout, MCPB Desktop acceptance and final cross-surface owner acceptance remain separate gates.

What's in 1.15.2

Kinocut 1.15.2 ships the verified sound line and an adversarial-hardened engine. Sound mastering now preserves stereo through the measured mastering chain with real caption-speech distance profiles, and sound-qa-asr fails closed without a real recognition request. Four guarded local Revideo operations join the public surface with rendered-format validation, and release-checkpoint passes the MCP release gate through from the CLI. Hardening from the adversarial audit: the MCP handshake reports the real kinocut version, a central guard rejects outputs that alias an operation's inputs, and oversized error echoes are truncated. Surface grows to 201 MCP tools / 173 CLI commands — the mcp-video==1.6.13 shim installs kinocut==1.15.2.

What's in 1.15.1

Kinocut 1.15.1 is a maintenance release on top of 1.15.0: the mcp-video compatibility shim 1.6.12 carries the registry ownership marker, the legacy io.github.KyaniteLabs/mcp-video registry entry was republished so directory listings (Glama, PulseMCP) point at living metadata (#469), and the object-matte engine now streams decode with scratch-directory caps — bounded memory on large inputs (#414). Surface stays 196 MCP / 167 CLI — no new public tool name.

What's in 1.15.0

Kinocut 1.15.0 (2026-08-19) was the Windows/diagnostics release: first-class Windows support and honest diagnostics on top of 1.14.x (same 360 dual-cam assembly and performance surface), driven by a community bug report. Surface stays 196 MCP / 167 CLI — no new public tool name.

• Honest MCP startup failures — kino --mcp keeps the "requires the 'mcp' package" hint only when mcp is genuinely absent; every other server-tree import failure now prints its real cause (filesystem paths redacted, markup-safe) and exits non-zero. Reported in #445.
• doctor verifies the MCP server import — new required mcp-server-import check imports the same server tree --mcp uses, so kino doctor can no longer report OK while MCP mode is broken.
• First-class Windows support — portable projectstore file locking (#446, landed with credit) makes kino --mcp importable on Windows; lock errors follow a contention-only contract; kino --help no longer crashes on legacy cp1252 consoles; and a windows-latest CI smoke job (lint, import checks, kino doctor, public-surface tests) guards the platform on every change.
• 360 dual-cam assembly — any stitched equirect MP4 (Insta360, Ricoh Theta, GoPro MAX, DJI Osmo 360, …) → reviewable 360_assembly_plan (desk / table / front_back, split / switch / PiP / single) → approve → FFmpeg v360 render. MCP: video_intent goal= + video_review_decide. Python: Client.propose_360_assembly / decide_360_assembly / render_360_assembly. Raw .insv / .360 rejected. Director is a plug (local first; cloud opt-in). Not an optimized-AI claim. Guide: docs/360_ASSEMBLY.md.
• Faster 360 and import path — single-pass split/PiP/switch filter_complex (source audio kept); sampled quality-gate analyze window; SHA-256 path/mtime cache; merge can skip re-probe when infos= is supplied; batched 360 storyboard stills; lazy mcp_video / CLI / Client.search_tools; doctor skips npx --yes unless Hyperframes is already on PATH.
• Lazy public import — import kinocut no longer eagerly loads Client/engines (PEP 562). from kinocut import Client and kinocut.Client is mcp_video.Client still hold.
• Ship-seam honesty — CLI/Client repurpose default 80; fail-closed unless ; candidate durable MCP applies before job success.

Also already on the published line from 1.13.x:

• Intent-verb surface — video_intent / intent (~10 verbs) to dry-run plans without silent media mutations
• Watching guardrail floor — video_review_run / video_review_decide (blackdetect/LUFS + first-15s inspection)
• B-roll proposals, ES-first caption translation, still/plate editor, cutfile-render, metric-qc
• Agent workflow engine, dedicated video rescue, layered compositing, Hyperframes, Shorts/Reels repurposing
• Counts at the 1.15.0 release — 196 MCP tools and 167 CLI commands

Full notes: CHANGELOG.md · published 1.16.1 on PyPI

Published quality and plain-file audio contracts are documented in quality evidence, audio mixing and projectstore lifecycle. See Unreleased for future local changes beyond the published package.

Beyond 1.16.1 (draft / gated)

The published compatibility installer mcp-video==1.6.16 pins kinocut==1.16.1. Fleet upgrade and running-process verification instructions are in docs/FLEET_UPGRADE.md.

The 2026-10-05 publication checkpoint is 1.16.1. Draft/gated work stays out of release claims until publication is verified. Live directory submissions and launch posts remain operator/human residual (docs/HUMAN_GATES.md) and are not claimed complete.

Staged and Gated Surfaces

While the core FFmpeg editing, 360 assembly, workflow engine, still/plate editing, AI-video review/salvage, and thin sound join are integrated on the published 1.16.x line, the following surfaces remain gated, partial, or unreleased:

• Desktop MCPB Packaging: The unsigned staged package (mcpb/) is labeled user-configured-local-access. It requires local Node/Python/Kinocut/FFmpeg, is not a sandbox, and is distinct from future native-runtime work. See docs/MCPB.md.
• Sonic World Audio (kinocut_sound): The shipped public boundary remains the thin S12 join; S13 packages and host joins exist as deeper internals. The historical July S14 receipt records two synthetic hardware classes, while the later August rerun records Apple Silicon and an explicit external_host_unavailable x86 residual. Its 64 clips of 0.15 seconds prove bounded plumbing, not a full episode and not human listening approval.
• Trusted Execution Kernel: The protected-timeline trusted execution kernel is post-program/gated and does not execute without the named upstream contract and human gating (docs/plans/2026-07-09-kinocut-trusted-execution-layer.md).
• Generative Plans and Local Caption Speech: Generative spend-cap and legacy TTS candidate records remain non-executable plans. An installed eSpeak NG engine can render real EN/ES stock caption speech through a hashed sound_voice_batch request and retain audio for mixing. Caption speech does not translate text or promise neural voice quality. No generation adapter exists; credentials or discovery alone do not establish execution.
• Product / object matte: Catalog and shop cutouts on the existing hyperframes-remove-background command (--model birefnet-general, optional extra kinocut[object-matte]). Default remains people. Not a new MCP/CLI name. Published in 1.16.1, first shipped in 1.15.1 (streaming decode + scratch caps, #414); the ONNX extra installs via kinocut[object-matte]. Guide: docs/PRODUCT_MATTE.md.

Product checklist: ROADMAP.md.

Agent Workflow Engine

Agents can plan, validate, render, recover, and prove a multi-step local video job from a single JSON job-spec — through MCP (video_workflow_*), the CLI (workflow-*), or the Python client (Client.workflow_*) — with receipts strong enough for another agent or a human to trust before and after a render. Ops are a small allowlist (probe | trim | resize | convert | crop | merge | add_text | composite_layers | burn_in) mapped 1:1 to the same vetted engine functions the individual tools use; media references are symbolic and workspace-confined; everything fails closed.

{
  "schema_version": 1,
  "name": "captioned-vertical-short",
  "sources": { "hero": { "path": "input/hero.mp4" } },
  "steps": [
    { "id": "trim-hero", "op": "trim", "inputs": { "src": "@sources.hero" },
      "params": { "start": 0, "duration": 6 }, "output": "@work/hero_trim.mp4" },
    { "id": "vertical", "op": "resize", "inputs": { "src": "@work/hero_trim.mp4" },
      "params": { "width": 1080, "height": 1920 }, "output": "@work/hero_vertical.mp4" },
    { "id": "caption", "op": "add_text", "inputs": { "src": "@work/hero_vertical.mp4" },
      "params": { "text": "Watch this", "position": "bottom-center" }, "output": "@outputs.master" }
  ],
  "outputs": { "master": { "path": "output/final.mp4" } }
}

kino workflow-validate --spec job.json    # cheap structural gate, no render
kino workflow-plan     --spec job.json --save-plan plan.json     # dry-run op graph + hashes
kino workflow-render   --spec job.json --save-receipt receipt.json   # execute + provenance receipt
kino workflow-inspect  --receipt receipt.json    # read-only integrity re-check

The render receipt records per-step input/output hashes, a resume cursor, and a cleanup manifest, all with workspace-relative paths:

{
  "receipt_kind": "workflow",
  "versions": { "mcp_video": "1.13.2", "ffmpeg": "8.1" },
  "spec_hash": "sha256:be2f3a9b...",
  "steps": [
    { "id": "trim-hero", "op": "trim", "status": "completed",
      "input_hashes": { "src": "sha256:3b976d49..." },
      "output": "work/be2f3a9b-2effedb3/mcp_video_hero_trim.mp4", "output_hash": "sha256:00727499..." },
    { "id": "caption", "op": "add_text", "status": "completed",
      "output": "output/final.mp4", "output_hash": "sha256:8633ad2a..." }
  ],
  "cleanup_manifest": { "cleaned": true, "policy": "clean-on-success" },
  "resume_cursor": { "last_completed_step": "caption", "next_step": null },
  "status": "completed",
  "render_determinism_scope": "spec/input/output hashes are deterministic; rendered bytes may vary across FFmpeg builds"
}

--all-variants emits N distinct outputs from one declaration, and --resume continues a job that failed with its intermediates kept (fail-closed on a changed spec). Full schema, @ref grammar, variants, resume, and cleanup are in docs/WORKFLOWS.md; a runnable spec is in examples/workflows/.

Governed AI-video review

In Kinocut, a contract-first path is provided for agent-edited media that must stay attributable and reviewable:

1. Ingest the source into a private content-addressed project (video_ingest / video-ingest)
2. Preflight + temporal inspection on the stored asset (video_preflight, video_inspect_temporal)
3. Verdict + acceptance with exact human evidence (video_verdict, video_acceptance_eval)
4. Bounded derivatives only — audio-preserving body swap or allowlisted salvage recipes (video_body_swap, video_salvage), each with lineage and a fresh non-approved review slot

There is no force/bypass flag. Analyzer output alone cannot approve. Stale, aliased, or protected inputs fail closed. Operating guide: docs/AI_VIDEO_REVIEW_AND_SALVAGE.md. These surfaces are fully integrated in the published 1.15.x line — see Status and releases.

Dedicated Video Rescue

For "fix this clip" requests where the story and timeline must remain unchanged, use the review-first rescue pipeline. Plan and inspect the diagnosis, approve only safe repair IDs, render, then inspect the verified package. The source stays immutable; master and universal sharing copy are always verified; optional captions remain sidecars. See docs/RESCUE.md for CLI, MCP, Python, cancellation, resume, and stable errors.

Layered Compositing

composite-layers / video_composite_layers adds a spec-driven ordered layer stack for agents that need more than two-shot overlay primitives. It supports image, video, and solid layers; normal alpha compositing; per-layer opacity; x/y placement; transform sizing; timing windows; and mask/matte alpha sources — plus full-canvas and positioned blend modes (multiply, screen, overlay, darken, lighten) and rotation with a new pivot reference point. Dry-run plans and deterministic layer_plan v2 receipts capture source, filtergraph, and output hashes.

kino composite-layers --spec layers.json --dry-run --save-layer-plan layer-plan.json
kino composite-layers --spec layers.json -o out.mp4 --save-layer-plan layer-plan.json

Use composite-layers when an agent needs a planned stack of overlays, mattes, lower thirds, blurback plates or platform variants reviewed before rendering. Non-normal blends support opacity and start/duration windows in two geometries: full-canvas at {0,0} without explicit sizing, or a positioned rectangle with both positive integer width and height and an integral nonnegative in-canvas position. RGB blending avoids applying color arithmetic to subsampled chroma planes. Scale, rotation/pivot, mask/matte, fractional positions and out-of-canvas rectangles remain deferred and fail closed with unsupported_blend_geometry. Video layers and masks begin playback at the declared layer start. Keep sources and masks inside the spec directory, placing the spec in a common parent of its media. Output remains video-only.

360 dual-cam assembly

Any stitched equirect 360 MP4 (Insta360, Ricoh Theta, GoPro MAX, DJI, …) can become a two-cam 16:9 or 9:16 edit without a new tool name. Export from the camera app first — Kinocut does not stitch .insv or GoPro .360.

from kinocut import Client

video = Client()
plan = video.propose_360_assembly("x4-export.mp4", goal="desk 360 split 9:16")
approved = video.decide_360_assembly(plan, "approve")
video.render_360_assembly(approved, "desk-split.mp4")

Agents: video_intent with a 360/desk/table goal and source=, then video_review_decide. Full contract: docs/360_ASSEMBLY.md.

Still / Image Editing

Kinocut treats multi-still packages as first-class media — plan → receipt → fail-closed gate, same safety posture as video rescue. Requires pip install "kinocut[image]"; run kino doctor to verify the image stack.

Workflow: establish a hero plate → edit beats toward it → match shared WB/exposure → grade (optional signal LUT) → cohesion gate.

# Image analysis: extract colors, generate palettes, analyze products
kino image-extract-colors product.jpg -n 8
kino image-generate-palette scene.png --harmony triadic
kino image-analyze-product listing.jpg --use-ai

# Still/plate editing: match package to hero, grade, gate
kino still-match --hero establish.png --inputs beat1.png beat2.png --output-dir out/matched
kino still-grade --inputs out/matched/*.png --hero establish.png --output-dir out/graded
kino still-gate  --inputs out/graded/*.png --output-dir out/gate

# Establish-locked edit with plan/receipt (dry-run first)
kino image-edit --source beat.png --reference establish.png \
  --intent "match establish world and light" --output-dir out/edit --dry-run

# Or run the full pipeline in one shot
kino still-package --establish establish.png --beats beat1.png beat2.png --output-dir out/pkg

from kinocut import Client

c = Client()
c.still_match(hero="establish.png", inputs=["a.png", "b.png"], output_dir="out/m")
c.still_gate(inputs=["out/m/a_matched.png", "out/m/b_matched.png"], output_dir="out/g")

still-gate fails closed on luma spread and shadow green/cyan metrics; every tool writes a JSON receipt with hashes and gains. Paid generative backends stay off by default (prefer=edit, allow_paid_gen=false). Full guide: docs/STILL_PLATES.md.

Public Discovery

Kinocut is built to be findable and citable by both search engines and AI answer engines:

• Canonical product URL: https://kinocut.dev/
• GitHub README + llms.txt with entity facts, install commands, and safety rules
• Official MCP Registry record under tech.kyanitelabs/kinocut
• FAQ answers in this README and docs/faq.md (answer-first, versioned claims)

Kinocut vs raw FFmpeg (and vs cloud editors)

Why It Exists

AI agents can write FFmpeg commands, but they should not have to guess flags, parse brittle stderr, or silently publish broken media. Kinocut gives agents typed operations, inspectable tool metadata, structured results, preflight guardrails, and quality checkpoints so a video workflow can be automated and reviewed without turning into shell-command roulette.

Use it when you want an AI assistant to:

• trim, merge, resize, crop, rotate, transcode, or export video;
• add text, subtitles, watermarks, overlays, filters, fades, effects, and transitions;
• extract audio, normalize audio, synthesize audio, add generated audio, or create waveforms;
• detect scenes, make thumbnails, generate storyboards, compare quality, and create release checkpoints;
• match, grade, and gate still packages with establish-locked color cohesion, or analyze product images for colors and palettes;
• scaffold cinematic projects, read STYLE_/NEG_ blocks, parse storyboard tables, and expand shot prompts;
• create new Hyperframes projects, inspect rendered layouts, capture websites, generate local speech, remove backgrounds, and post-process the result with FFmpeg tools;
• repurpose one source video into vertical, horizontal, and square local delivery packages with manifests and review artifacts;
• drive repeatable media workflows from Claude Code, Cursor, Codex-style clients, scripts, or CI.

Installation

Prerequisite: FFmpeg must be installed and available on PATH.

# macOS
brew install ffmpeg

# Ubuntu/Debian
sudo apt install ffmpeg

Run without a global install:

uvx --from kinocut kino doctor

Or install with pip:

pip install kinocut
kino doctor

For Claude Desktop-style MCPB installs, Kinocut includes a staged local package at mcpb/ and a local build script:

python3 scripts/build-mcpb.py

This unsigned user-configured-local-access package launches an existing Python environment with the exact Kinocut version and still requires local Node, FFmpeg, and ffprobe. It is not a sandbox. Native self-contained bundles remain separate future work. See docs/MCPB.md.

Optional C2PA signing for final MP4 exports is available when c2patool and a manifest/signer are configured. Signing is off by default and only reports signed after a verification read succeeds. See docs/C2PA_PROVENANCE.md.

Hyperframes tools additionally need Node.js 22+ and a resolvable Hyperframes CLI. Install/pin Hyperframes in the active Node package layout, add hyperframes to PATH, or set MCP_VIDEO_HYPERFRAMES_COMMAND.

Which extra do I need?

The core install covers all FFmpeg editing tools. Optional features ship as extras — install only what you use:

Mix freely, e.g. pip install "kinocut[transcribe,image]". Run kino doctor afterward — it reports exactly which features are available and what is missing.

Upgrading from mcp-video

Kinocut preserves the original surface during the rename window. Existing installs can upgrade without changing code:

pip install --upgrade 'mcp-video==1.6.16'
pip install --force-reinstall --no-deps 'kinocut==1.16.1'
mcp-video doctor

Older shim installer records can delete the shared executable during an upgrade, even when dependency checks pass. The canonical reinstall restores ownership; use the intended release pin and preserve configured extras in locked deployments. See fleet migration instructions for the pinned sequence.

Published mcp-video==1.6.16 is a metadata-only compatibility installer for kinocut==1.16.1. The mcp_video import, mcp-video command, MCP_VIDEO_* environment variables, ~/.mcp-video data directory, mcp-video:// resource URIs, and existing receipt keys remain supported on the 1.14.x+ line. New integrations should use kinocut, from kinocut import Client, and the kino command.

En español

Kinocut es un servidor MCP de edición de video para agentes de IA. La última versión publicada es 1.16.1 (pip install kinocut, 203 herramientas MCP / 177 CLI). Incluye ensamblaje 360 de dual-cam desde un MP4 equirectangular ya stitched (no .insv), import perezoso PEP 562 y el mismo surface FFmpeg tipado para recortar, unir, subtitular, mezclar audio, efectos y reutilizar contenido (Shorts, Reels, TikTok), motor de flujos (workflow) con recibos verificables, rescate de video, revisión AI-video gobernada y barreras de seguridad antes de renderizar. Programas humanos residuales (directorios, lanzamiento) no se reclaman completos.

Requisito: FFmpeg instalado y disponible en el PATH.

# macOS
brew install ffmpeg

# Ubuntu/Debian
sudo apt install ffmpeg

# Instalación y diagnóstico
pip install kinocut
kino doctor

Para Claude Code:

claude mcp add kinocut -- uvx --from kinocut kino --mcp

kino doctor informa qué funciones están disponibles y qué falta instalar. La documentación completa está en inglés; los mensajes de error principales son bilingües.

Quick Start

Golden path (60 seconds)

Prove the install works before wiring an agent host:

pip install -e .          # or: pip install kinocut
kino doctor               # required checks must pass
python scripts/golden_path.py

Success criteria and failure recovery: docs/GOLDEN_PATH.md.
Shareable pack (receipt + quality + media): python scripts/generate_golden_pack.py → demo/golden-pack/.

Try the receipt-backed proof first

From a clone of this repo, run the smallest confidence workflow before wiring an agent host:

uv run --no-project --with kinocut python workflows/05-confidence-baseline/workflow.py
uv run --no-project --with kinocut python workflows/benchmarks/run_confidence_benchmark.py

The workflow generates a tiny source clip, creates a checked vertical video, runs quality/release checkpoint steps, and writes workflows/05-confidence-baseline/output/video_receipt.json.

Proof notes live in docs/proofs/. Public marketing claims (version, tool counts, URLs) live in docs/public_claims.json and are CI-guarded.

Claude Code

claude mcp add kinocut -- uvx --from kinocut kino --mcp

Claude Desktop

{
  "mcpServers": {
    "kinocut": {
      "command": "uvx",
      "args": ["--from", "kinocut", "kino", "--mcp"]
    }
  }
}

Cursor

{
  "mcpServers": {
    "kinocut": {
      "command": "uvx",
      "args": ["--from", "kinocut", "kino", "--mcp"]
    }
  }
}

Then ask your agent:

Trim this interview into a 45-second vertical clip, add burned captions, normalize the audio, make a thumbnail, and create a release checkpoint before export.

Agent Skill

Kinocut includes a public agent skill at skills/kinocut/SKILL.md. Use $kinocut in compatible agent hosts when you want the agent to choose between the MCP server, CLI, and Python client while preserving the inspect, edit, verify, and human-review workflow.

For path-based short-form packages from current tools only (no invented commands, no external publish), see skills/kinocut-repurpose/SKILL.md. That skill is an explicit marketing seed; the durable kernel-backed repurposing product is still on the trusted-execution roadmap.

Python Client

from kinocut import Client

editor = Client()

clip = editor.trim("interview.mp4", start="00:02:15", duration="00:00:45")
caption_file = "captions.srt"
editor.ai_transcribe(clip.output_path, output_srt=caption_file)
captioned = editor.subtitles(clip.output_path, subtitle_file=caption_file)
vertical = editor.resize(captioned.output_path, aspect_ratio="9:16")
checkpoint = editor.release_checkpoint(vertical.output_path)

print(checkpoint["thumbnail"])
print(checkpoint["storyboard"])

CLI

kino info interview.mp4
kino trim interview.mp4 -s 00:02:15 -d 45
kino video-ai-transcribe clip.mp4 --output captions.srt
kino subtitles clip.mp4 captions.srt
kino resize clip.mp4 --aspect-ratio 9:16
kino video-quality-check clip.mp4
kino repurpose clip.mp4 --platforms youtube-shorts instagram-reel tiktok
kino image-extract-colors product.jpg
kino still-package --establish hero.png --beats shot1.png shot2.png --output-dir out/stills

What Agents Can Do

MCP Tools

Published 1.16.1 registers 203 MCP tools and 177 CLI commands (the matching development checkout has the same surface at release). The table summarizes core categories — search_tools discovers the exact operation without loading every description.

from kinocut import Client

editor = Client()
matches = editor.search_tools("subtitle")
print(matches["tools"])

Full reference: docs/TOOLS.md

Agent-Safe Workflow

For autonomous agents, the intended path is inspect, edit, verify, then ask a human to review release artifacts:

from kinocut import Client

client = Client()

print(client.inspect("trim"))

result = client.pipeline(
    [
        {"op": "trim", "input": "source.mp4", "start": "00:01:00", "duration": "00:00:45"},
        {"op": "add_text", "text": "Launch clip", "position": "top-center"},
        {"op": "normalize_audio"},
        {"op": "resize", "aspect_ratio": "9:16"},
        {"op": "export", "quality": "high"},
        {"op": "release_checkpoint"},
    ],
    output_path="final-short.mp4",
)

Safety contract:

• Media-producing calls return structured results with output paths.
• High-risk edit paths now run preflight guardrails before FFmpeg execution: filter bounds, merge compatibility, audio mix volume/timing, overlay/watermark/chroma opacity and similarity, animated text timing/overflow, and grid/split-screen mismatch warnings.
• Analysis and discovery calls return structured JSON reports.
• Tool discovery is available through search_tools() and Client.inspect().
• Unexpected keyword errors are converted into actionable MCPVideoError guidance.
• Do not publish agent-generated video without video_quality_check, video_release_checkpoint, and human visual/audio inspection.
• For governed AI-video derivatives, require stored identities, active human decision evidence, and a fresh review slot after every salvage or body-swap — never raw FFmpeg workarounds labeled as governed.

Changelog

1.16.1 (2026-10-05):

• Maintenance fixes: macOS render cleanup (live process supervisor, zombie-tolerant quiescence) and MOV/MP4 faststart preservation for descriptor-anchored outputs.
• C2PA signing and object-matte/FSRCNN model downloads now publish through owned staging files only after verification.
• FFmpeg 8.1 filter discovery; source-audio preservation in GPU shader and Real-ESRGAN reconstruction; typed errors for shader/upscale probe failures.
• Surface stays 203 MCP / 177 CLI; mcp-video==1.6.16 installs kinocut==1.16.1.

1.16.0 (2026-10-02):

• Operator-route parity across MCP/CLI/Python (audio mixing, ducking/HLS, motion receipts, 360 intent/review controls, operation estimates); persisted MCP repurpose quality/checkpoint policy.
• Bounded input hardening (CLI plan/review JSON size and nesting caps, explicit JSON booleans, duplicate-key rejection).
• Added Client.record_motion_acceptance, temporal motion_coherence evidence, opt-in Anthropic keyframe review, structured FFmpeg dependency advisories, and Gemini CLI setup examples.
• Surface grows to 203 MCP / 177 CLI; mcp-video==1.6.15 installs kinocut==1.16.0.

1.15.3 (2026-09-25):

• Publication release: PyPI and npm verified 2026-09-30; mcp-video==1.6.14 installs kinocut==1.15.3. Surface stays 201 MCP / 173 CLI.

1.15.2 (2026-09-24):

• Verified stereo sound mastering through the measured mastering chain (#523/#525/#526); real caption speech processed with distance profiles; sound-qa-asr fails closed without a real recognition request (#432).
• Guarded local Revideo operations exposed on the public surface with rendered-format validation (#508); release-checkpoint passes through the MCP release gate from the CLI.
• Adversarial hardening: the MCP handshake reports the real kinocut version (F1), a central self-overwrite guard rejects outputs aliasing inputs (F2), error echoes truncate oversized values (F4).
• Absolute trim-end semantics preserved (#493); Windows drive paths keep their prefix (#496); empty/blank ASS caption tracks fail closed instead of rendering empty output.
• Surface grows to 201 MCP / 173 CLI; mcp-video==1.6.13 installs kinocut==1.15.2.

1.15.1 (2026-08-31):

• Registry ownership: mcp-video shim 1.6.12 + legacy registry republish (#469).
• Object-matte streaming decode with scratch caps (#414).
• mcp-video==1.6.12 installs kinocut==1.15.1.

1.15.0 (2026-08-19):

• Honest MCP startup failures — kino --mcp reports real import causes; paths redacted, markup-safe (#448, from #445).
• doctor mcp-server-import check — required core check; doctor can't say OK while --mcp is broken (#449).
• First-class Windows support — portable projectstore file locking (#446), contention-only lock contract, UTF-8 stdio, windows-latest smoke job.
• Community-driven release — @gerardoscaglia-creator credited in Contributing and the CHANGELOG acknowledgements.
• mcp-video==1.6.11 installs kinocut==1.15.0.

1.14.1 (2026-08-13):

• Patch on 1.14.0: pyright NoReturn on rescue cancel so GitHub Lint stays green. Same 360 + perf surface.
• mcp-video==1.6.10 installs kinocut==1.14.1.

1.14.0 (2026-08-13):

• Published 360 dual-cam assembly from a stitched equirect MP4 (camera-agnostic; not X4-only) via video_intent / Client.propose_360_assembly — still 196 MCP / 167 CLI.
• Landed the performance committee top-10: single-pass 360 filter_complex (audio kept), sampled QC, lazy imports, SHA cache, batched storyboard stills.
• mcp-video==1.6.9 installs kinocut==1.14.0.

1.13.0 (2026-08-07):

• Added Intent-verb surface (video_intent / intent), Watching guardrail floor (video_review_run / video_review_decide), B-roll proposals (video_propose_broll), Caption translation ES-first (video_translate_captions), and Trusted execution (TE) quality-of-life additions.
• Expanded published surface to 196 MCP tools and 167 CLI commands (fully synchronized across standard, test, and client interfaces).
• Hardened path validation for caption translation, brand kit, and OTIO exports.

1.12.0 (2026-08-07):

• Added Still/plate editor surface (still-match, still-grade, still-gate, image-edit, still-package) with fail-closed cohesion gating.

1.11.x (2026-07-24):

• Added thin kinocut_sound S12 public join adapters (sound-capabilities, sound-plan-validate, sound-mix-render, etc.).
• Post-theme security hardening to prevent absolute host-path leaks and ensure all discovered operations have fail-closed invoke paths.

See CHANGELOG.md for full historical release notes, or view the GitHub Releases page.

FAQ

What is Kinocut?

Kinocut is a free, open-source MCP server, Python library, and kino CLI for AI-agent video editing. It wraps FFmpeg (and optional Hyperframes/Whisper extras) with preflight guardrails, Video Receipts, and quality checkpoints. It was formerly named mcp-video.

How do I install it?

brew install ffmpeg   # or apt install ffmpeg
pip install kinocut
kino doctor
claude mcp add kinocut -- uvx --from kinocut kino --mcp

Is it free and local-first?

Yes. Apache-2.0, runs on your machine, no Kinocut account or API key required for the core surface, and media is not uploaded to a Kinocut cloud.

Local compute and storage still have costs. Optional providers and the host agent can charge separately, and prompts or tool results can reach the host's model provider. Local execution alone does not establish zero data egress.

Which agents work with it?

Any MCP-compatible client that can run a local stdio server (Claude Code, Cursor, Windsurf, Cline, and similar). You can also use the Python client or CLI without an agent.

How many tools are there?

Published 1.16.1 provides 203 MCP tools / 177 CLI commands. This branch has the same registered surface at release. Query the installed server's tool schemas for exact parameters. 360 assembly still reuses video_intent and video_review_decide.

Can Kinocut edit Insta360 X4 360 video?

Yes — from a stitched 360 MP4, not .insv. Propose a 360_assembly_plan, approve it, then render split/switch/PiP/single. Works for other stitched equirect cameras too. See docs/360_ASSEMBLY.md.

Was it called mcp-video?

Yes. The current published mcp-video==1.6.16 installs kinocut==1.16.1; the historical mcp-video==1.6.11 installed kinocut==1.15.0. Compatibility imports, CLI name, env vars, data dir, resource URIs, and receipt keys remain supported on the 1.14.x+ line.

Can Kinocut turn an Insta360 X4 file into a two-cam edit?

Yes: export a stitched 360 MP4, then propose → approve → render. .insv is rejected. This shipped in 1.14.1 and remains in published 1.16.1.

How is Kinocut different?

Unlike raw FFmpeg scripts or unguarded agent shells, Kinocut validates tools and emits receipts.

Is Kinocut production software?

Treat the README status and release tags as source of truth for maturity. Validate against your own requirements before production use.

More answers: docs/faq.md · on-site FAQ: kinocut.dev/#faq

Documentation

• Documentation map
• Product site
• Tool reference
• Python client reference
• CLI reference
• Agent workflow engine
• Video receipts
• Video rescue
• 360 dual-cam assembly
• Post-rescue planning
• AI-video review and salvage
• AI-video contracts
• AI-video inspection
• C2PA provenance
• MCPB packaging
• Product roadmap
• Trusted execution layer plan
• Wishlist / 1.8 program parallel plan
• Sound program status
• AI agent discovery guide

Testing

Development verification lives in docs/TESTING.md. Keep public-surface, media workflow, and security checks current when changing tool behavior.

Development

git clone https://github.com/KyaniteLabs/kinocut.git
cd kinocut
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest tests/ -v -m "not slow and not hyperframes"

Community

• Contributing
• Code of Conduct
• Governance
• Maintainers
• Security
• Support
• Roadmap
• Changelog
• GitHub issues

Community deploys

Operators who want a public MCP URL without opening home-router ports can study the community Apache-2.0 reference stack that puts Kinocut + Hyperframes behind Tailscale Funnel, Caddy, and Google OAuth (mcp-auth-proxy): hyperframes-selfhost (not affiliated with KyaniteLabs; evaluate security for your threat model).

License

Apache 2.0. See LICENSE.

Built with FFmpeg, Hyperframes, and the Model Context Protocol.

Part of KyaniteLabs

More from KyaniteLabs. Related projects:

• Epoch — time-estimation MCP server (PERT) for AI agents
• DialectOS — Spanish dialect localization MCP server & CLI
• checkyourself — local-first production-readiness checks for AI-built code

→ More at kyanitelabs.tech

If Kinocut is useful to you, star or watch it on GitHub — it helps other agent builders find it. If it saved you real editing hours: ko-fi.com/kyanitelabs.

Built by Simon Gonzalez De Cruz — available for Forward-Deployed / Applied-AI engineering and contract work via the public profile links above.

Who it's for

• Primary: AI agent builders, Claude Code/Cursor users, and local media operators
• Use when you need to edit, caption, repurpose, and quality-gate video with typed FFmpeg tools
• Skip if you need a hosted cloud editor or untyped FFmpeg shell

Status

• Maintained as of 2026 on the default branch
• Prefer release tags when pinning dependencies
• Report issues on the canonical GitHub repository

Agent surface

• Coding agents: read this README first, then repo docs/AGENTS.md if present
• Prefer machine-readable briefs (llms.txt) when the repo ships one
• MCP or skill entrypoints are documented in-repo when applicable

Contributing

Issues and PRs are welcome on GitHub. Keep public docs free of secrets and machine-local paths.

Contributors

Kinocut improves in public, and outside contributions shape it — 1.15.0's diagnostics fixes trace directly to one excellent bug report. Reports with reproduction steps and a real traceback are contributions, and contributors are credited by name in the CHANGELOG acknowledgements and the release notes.

• @gerardoscaglia-creator — his Windows MCP-mode investigation (#445) exposed both diagnostics gaps fixed in 1.15.0, and his portable file-locking fix (#446) landed in 1.15.0 with credit. Thank you.
• @pedropaav-art — Windows filter-option path escaping (#458), merged 2026-08-18, shipped in 1.15.0.
• @linhaixin45-cmyk — kino doctor Python-version floor with a visible install hint (#459), merged 2026-08-18, shipped in 1.15.0.
• @WohaibHasan — absolute trim-end handling (#493), Windows drive preservation (#496), and portable UTF-8 CLI test capture (#498).
• @betsmayank — first merged external fix: Hyperframes init no longer hangs under MCP without a TTY (#361).
• @austinwmson and @ismailkattakath — external reports that closed real gaps: remotion_still runtime props (#306) and the self-host Funnel + OAuth reference stack (#432).
• @Dodothereal and @OyaAIProd (account no longer exists) — early external PRs (ASS PlayRes for vertical burns #378, SafeSkill badges).

Cross-platform architecture and current support evidence: platform parity. Web/mobile shared processing and native clients are not implemented in this release.