vibeedit

VibeEdit

VibeEdit is a frame-accurate video-production library for AI coding harnesses. It provides a VibeEdit-owned Python media API, a deterministic JavaScript/HTML motion runtime, one shared CompositionSpec, a searchable local catalog, safe skill installation, rendering, and output verification.

This source targets qualification candidate Python 0.2.0b2 / npm 0.2.0-beta.2. It adds portable projects and corrections, focused local review tools, explicit model packs, source-acquisition receipts and an optional Pi companion. It is not yet qualified as the complete launch. Use its exact release manifest to establish archive availability and source identity; this source checkout does not establish publication. Registry installation and complete customer journeys are separate gates. The previous public v0.2.0-beta.1 qualification prerelease remains pinned to 7ff8ff1a9caf1f53d7f8cba7fbbfec97d0dd61e1 with its original archives intact; it does not contain the beta.2 update fixes.

VibeEdit-owned code is MIT licensed. Optional models and fonts have their own terms in THIRD_PARTY_NOTICES.md. Myriograph owns the three paid intelligence services and metered companion access. Ordinary local editing requires no account or Pi companion.

Install

Give your agent the version-pinned setup skill: "Read this VibeEdit setup skill, install it for this project, verify it, and help me make a video."

The agent verifies the release manifest, checksums and matching skill before installing exact archives into isolated environments. If that version's manifest is unavailable, stop and report that the candidate has not been published; the previous beta.1 setup is a separately pinned fallback. Do not mix their Python and Node versions or assume registry availability. For manual installation, download and verify the files from the same release; the commands below use those verified local files.

Lightweight Python CLI and media orchestration:

uv tool install ./vibeedit-0.2.0b2-py3-none-any.whl
vibeedit doctor

For a manual installation inside a Python virtual environment, choose only the needed extras below. These pip commands do not modify an existing uv tool environment; the setup skill handles isolated optional-runtime installation.

pip install "./vibeedit-0.2.0b2-py3-none-any.whl[effects]"
pip install "./vibeedit-0.2.0b2-py3-none-any.whl[browser]"
pip install "./vibeedit-0.2.0b2-py3-none-any.whl[vision]"
pip install "./vibeedit-0.2.0b2-py3-none-any.whl[sam]"
pip install "./vibeedit-0.2.0b2-py3-none-any.whl[all]"
vibeedit setup --effects
vibeedit setup --browser
vibeedit setup --sam
vibeedit doctor --json

The effects extra activates 333 reviewed deterministic frame presets: 200 filters, 112 effects, and 21 transitions. Every preset has a stable catalog ID, bounded parameters, source hashes, and an exhaustive execution result in catalog/preset-validation.json.

from vibeedit import apply_media_preset, render_transition_preset

graded = apply_media_preset(
    rgba_frame,
    "vibeedit://effect/filters-cinematic-teal-orange",
    parameter_overrides={"intensity": 0.42},
    progress=0.4,
)
transition = render_transition_preset(
    graded,
    next_rgba_frame,
    "vibeedit://transition/transitions-core-film-burn",
    progress=0.5,
)

The vision extra exposes real OpenCV face detection, person/body detection, and deterministic temporal face tracking through CapabilityRouter. Tracking writes a normalized JSON artifact and returns a public TrackingArtifact with source, runtime, and cache-key provenance. On macOS, vibeedit setup --vision builds the packaged VibeEdit-owned Swift runner and enables its explicitly declared face, body, and pose requests. The same explicit setup downloads a checksum-pinned 29.5 MB SSD-MobileNetV1 ONNX model for portable general-object detection where ONNX Runtime publishes a wheel. ONNX Runtime 1.27 has no macOS Intel wheel, so that platform retains OpenCV face/body tracking and Apple Vision while reporting the portable object provider as unsupported. Unsupported providers remain explicit in doctor; they are never inferred merely from the operating system.

vibeedit setup --sam is an explicit 211.7 MB optional download. It installs a checksum-pinned official SAM 2.1 source revision and tiny checkpoint in the VibeEdit cache; the source and weights are Apache-2.0. The runner uses Torch/CUDA, Torch/MPS, or CPU in that order and records the model version, checkpoint digest, prompt, source identity, actual device/runtime versions, and cache key. A clean-wheel Torch/MPS proof with controlled and natural masks is recorded in the historical SAM2.1 proof for an earlier release. The new model-pack command supports pinned TransNet, SAM 2.1 and SAM 3.1 MLX setup. Fresh inference qualification for this candidate is pending; setup availability is not model accuracy evidence. No model weight is part of the base wheel or npm archive.

Node SDK, HTML motion runtime, and Node-facing CLI:

npm install ./vibeedit-0.2.0-beta.2.tgz
npx vibeedit doctor --json

The Node package owns CompositionSpec, catalog, skills, and HTML motion APIs. Media rendering commands bridge to an installed Python VibeEdit package. Run npx vibeedit doctor --json to see both capabilities and exact next actions.

Agent-package guides: projects/review, models and reviewed effects, Pi workers/goals, Myriograph funding/services, and operations/rollback. Launch state keeps actual qualification separate from supported command contracts.

Agent quick start

Use this sequence when an agent has no prior VibeEdit context:

vibeedit doctor --json
vibeedit examples list --details --json
vibeedit catalog search kinetic --compact --limit 5 --json
vibeedit examples create basic-generated ./vibeedit-work --json
vibeedit render ./vibeedit-work/basic-generated/composition.json --json

ready: true means the core FFmpeg renderer is ready. It does not claim every optional model or browser capability is installed; inspect the readiness and capabilities fields for those distinctions. Compact catalog search is intended for low-token agent discovery. Omit --compact when the full parameter, provenance, prompt, and validation contract is needed.

Compact results are ranked deterministically and include the selected stable ID, intent, required capability, backend, determinism, parameter count, platform compatibility, setup cost, confidence, and reason. Search does not load skill bodies; inspect only the selected result when more detail is needed.

Before a human revision, inspect the dependency-aware invalidation plan:

from vibeedit import plan_revision, render_revision

plan = plan_revision(previous_spec, revised_spec)
print(plan["dirtyFrameRanges"], plan["expectedReuse"])
render_revision(previous_spec, revised_spec, "previous.mp4", "revised.mkv")

The same preflight is available without Python code:

vibeedit revision plan previous.json revised.json --json

For bounded browser-text changes with caching enabled, VibeEdit reuses content-addressed composite frames outside the changed layer's placement. The render provenance sidecar reports work.framesRendered, work.framesReused, and the reuse kind. Compatible container-only changes use stream-copy remuxing without decoding video. Audio-clip and procedural-SFX parameter revisions remix audio and stream-copy the encoded video. Bounded browser-text changes still encode the complete final frame sequence; transition-range replacement is planned but not yet claimed as executable. A conservative scene-tail removal can stream-copy an exact number of approved prior video frames when all remaining visual layers are unchanged and every removed visual layer begins at or after the new end. If explicit music or SFX remain, VibeEdit rebuilds the revised audio mix instead of cutting AAC packets, then verifies the exact video frame count and prior artifact provenance. Mid-scene removal remains planned.

Install the Python beta directly from its GitHub release asset:

uv tool install https://github.com/Momenta-App/vibeedit-sdk/releases/download/v0.2.0-beta.2/vibeedit-0.2.0b2-py3-none-any.whl

Install the Node beta tarball in a project:

npm install https://github.com/Momenta-App/vibeedit-sdk/releases/download/v0.2.0-beta.2/vibeedit-0.2.0-beta.2.tgz

When downloading an existing release, use the checksum manifest distributed with that exact release:

shasum -a 256 -c SHA256SUMS

The build workflow emits future flat release manifests as SHA256SUMS.

Or clone and install the public source:

git clone https://github.com/Momenta-App/vibeedit-sdk.git
cd vibeedit-sdk
uv tool install .

From a Node project:

npm install github:Momenta-App/vibeedit-sdk

Or build and install the exact local artifacts:

uv build --out-dir dist/python
uv tool install dist/python/vibeedit-0.2.0b2-py3-none-any.whl
npm pack --pack-destination dist/npm
npm install ./dist/npm/vibeedit-0.2.0-beta.2.tgz

setup performs only explicitly requested work. It installs pinned browser and vision wheels into the current VibeEdit tool environment when missing. Browser setup then installs Playwright's pinned Chromium revision, records the resolved executable and SHA-256, and writes its manifest to the VibeEdit cache. On macOS, vision setup compiles the bundled Swift source with Apple Vision, records the binary SHA-256 and declared capability list, then downloads the exact portable object model declared in runtime-models/manifest.json. SAM setup downloads only its separately declared exact URLs, sizes, and SHA-256 values. No runtime model is included in the base wheel or npm archive.

Accelerated ordinary browser compositions

The Python package accepts normal bundled HTML/CSS/JS with local media. Chromium owns fonts, layout, graphics and composition. No CompositionSpec is needed for this workflow. On macOS 15+ ARM64, the CEF backend captures GPU surfaces, converts on Metal, and encodes opaque H.264 with VideoToolbox.

Install an exact qualified release wheel with the browser extra. Native setup requires FFmpeg, Xcode command-line tools, CMake, Ninja and Rust; it explicitly downloads the checksum-pinned CEF runtime and builds the bundled sources into the user cache:

vibeedit setup --cef --json
vibeedit examples create browser-composition ./projects
vibeedit render-browser ./projects/browser-composition/job.json --output ./movie.mp4 --backend cef --json

For portable Windows/Linux/macOS screenshot rendering:

vibeedit setup --browser --json
vibeedit render-browser ./projects/browser-composition/job.json --output ./movie.mp4 --backend screenshot --json

vibeedit browser-worker --backend cef accepts JSON lines containing id, job, and output to reuse a warm background browser. Python callers can import CEFService and render_browser_job from vibeedit.browser_render. The Node CLI delegates these commands to an installed Python package. Rendering does not download dependencies. Alpha exports use the screenshot fallback; native acceleration is currently macOS ARM64 only. See the renderer guide for media preparation, timing adapters and performance evidence.

One composition contract

All public APIs use schema/composition.schema.json. Timeline values are integer frames and frame rates are rational numbers.

from vibeedit import Canvas, Composition, FrameRate, MotionComponent, Placement, Track

composition = Composition(
    "title-card",
    Canvas(1920, 1080, FrameRate(30000, 1001)),
    90,
)
motion = composition.timeline.add_track(Track("M1", "motion", 10))
motion.add(
    MotionComponent(
        "title",
        Placement(0, 90),
        "vibeedit://text/negative",
        {"text": "NO EXCUSES"},
    )
)
composition.write("composition.json")

import { Canvas, Composition, FrameRate, MotionComponent, Placement, Timeline, Track } from "vibeedit";

const timeline = new Timeline();
const motion = timeline.addTrack(new Track({ id: "M1", kind: "motion", order: 10 }));
motion.add(new MotionComponent({
  id: "title",
  placement: new Placement(0, 90),
  componentId: "vibeedit://text/negative",
  props: { text: "NO EXCUSES" },
}));
const composition = new Composition({
  id: "title-card",
  canvas: new Canvas({ width: 1920, height: 1080, frameRate: new FrameRate(30000, 1001) }),
  durationFrames: 90,
  timeline,
});
composition.validate();

MoviePy, FFmpeg, Playwright, OpenCV, ONNX, Torch, MLX, and platform APIs are backend details. Their native objects never enter the public CompositionSpec.

Agents may supply raw HTML/CSS fragments or full documents through vibeedit://motion/html-css. This preferred contract loads local fonts/assets, automatically seeks CSS animations, rejects authored JavaScript, and includes a composable Motion Atoms vocabulary. The broader executable path accepts an inline HTML/CSS/JavaScript fragment or a complete locally bundled React, Vue, Svelte, Three.js, PixiJS, Canvas, WebGL, or WebGPU project through vibeedit://motion/html and vibeedit://motion/web-project. The persistent Chromium renderer loads the project once, waits for bundled fonts, and seeks it by integer frame. Run vibeedit inspect composition.json --json to see routing and library details. The complete authoring and deterministic-seek contract is documented in docs/HTML_MOTION_RUNTIME.md.

CLI

vibeedit init
vibeedit setup
vibeedit doctor
vibeedit inspect
vibeedit catalog list|search|open
vibeedit motion atoms
vibeedit examples list|create
vibeedit skills list|install|check|update|remove
vibeedit validate
vibeedit preview
vibeedit render
vibeedit verify
vibeedit clean
vibeedit mcp

Commands return actionable errors and support --json where automation benefits. Run a verified example with:

vibeedit examples create effect-transition
python effect-transition/render.py

vibeedit examples create mixed-python-html
python mixed-python-html/render.py

vibeedit examples create fan-edit
python fan-edit/render.py

vibeedit examples create face-follow-text
python face-follow-text/render.py

The local catalog stays in the background by default:

vibeedit catalog open --json           # resolve its packaged path; no tab
vibeedit catalog open --browser        # explicitly open a visible browser tab

The same example inventory is available without shell orchestration:

from vibeedit import create_example, render_example

directory = create_example("fan-edit")
render_example(directory)

import { createExample } from "vibeedit";

createExample("portable-motion-showcase");

Sixteen packaged examples cover generated assembly, effects/transitions, talking-head captions, portable kinetic typography, mixed media/HTML, a fan edit, beat analysis, layered sound design, a generated elliptical mask effect, tracking-driven face-follow text, multiple transitions, a VP9 alpha overlay, and conditional SAM 2.1 candidate review. Generated examples demonstrate rendering and composition; they do not establish real footage segmentation or editorial quality. The SAM example stops for candidate review and writes actionable evidence when its runtime is unavailable. Actual SAM 2.1 and SAM 3.1 MLX model runs, persisted mask corrections, and reviewed subject-effect renders use the separate vibeedit models workflow; their inference qualification remains open.

Catalog and skills

All 445 shipped capabilities have a stable vibeedit:// identifier, parameter schema, platforms/backends, provenance, validation evidence, example code, agent prompts, and an explicit preview state.

vibeedit catalog search stutter --json
vibeedit catalog open                 # print the local catalog path; no browser tab
vibeedit catalog open --browser       # explicitly open the catalog in a browser
vibeedit skills install vibeedit-workspace --harness codex

Skill install/update/remove operations track versions and checksums. They refuse to overwrite or remove user-modified skill files. The bundle contains 44 public skills adapted for the portable SDK from pinned canonical Git sources and records 23 rejected source skill packages with quarantine reasons in skills/migration-report.json. Packaging metadata remains outside the cloned skill directories. Each adaptation records its original source identity and the checksum of the adapted package.

Local tool server

vibeedit mcp starts a local stdio JSON-RPC/MCP-compatible adapter exposing catalog search, media inspection, composition creation, effects, transitions, motion, SFX placement, render, and verification. It calls the same library used by the CLI; it is not a separate implementation.

Reproducibility and security

Rendering uses integer-frame plans, normalized constant-frame-rate inputs, deterministic seeded effects, explicit runtime capability checks, safe subprocess argument arrays, content-based cache keys, and output probes. HTML components run locally with no network requirement. Catalog data is never executed as code.

See docs/SECURITY.md, docs/ARCHITECTURE.md, SBOM.spdx.json, and THIRD_PARTY_NOTICES.md for boundaries and inventories.

License

VibeEdit's owned package code, skills and templates use the MIT license in LICENSE.md. Fonts, optional model weights and third-party runtimes retain their own terms, recorded in THIRD_PARTY_NOTICES.md and SBOM.spdx.json. Myriograph is a separately metered service; package installation includes no paid inference credit. Earlier release legal records are historical evidence.