AviUtl2 Project API

Python API for manipulating AviUtl ver.2 project files (.aup2).

Overview

AviUtl ver.2 uses a text-based project format (.aup2) similar to INI files. This library provides:

• Parser: Read .aup2 files into Python objects
• Serializer: Write Python objects back to .aup2 format
• JSON Conversion: Export/import as JSON for LLM processing
• Validation: Timeline collision detection and frame calculations
• CLI Tool: Command-line interface for AI agent automation
• Preset System: Save and reuse animation/effect combinations
• Frame Preview: Render frames to PNG for Vision AI verification
• Smart Automation: Auto frame range, layer selection, and media duration detection
• Live Bridge (experimental): Connect to the currently open AviUtl2 project over a local Windows named pipe

What's New in 0.10.0

0.10.0 completes the 2026-09-19 official SDK mirror adoption with scene management, project files, file export, and MCP Phase 1, all gated on AviUtl2 2.1.10 (sdk_scene_crud host flag) and failing closed on older hosts.

• Scene CRUD (scene.list / scene.create / scene.switch) on 2.1.10+, plus scene background color in scene.get_current / scene.update_current. Scene delete and duplicate stay unsupported (no official SDK API).
• Project files (project.create / project.open / project.save) with explicit show_confirm=false automation handling.
• File export (export.start): async export through a host output plugin (e.g. MP4 Exporter (by えすご/Esugo)); the plugin name is caller-supplied because the SDK has no enumeration API. Completion detection uses the sequenced edit_state_changed event stream (null payload) combined with edit_state polling and an output-file existence probe.
• Object flags (object.flag.get / object.flag.set) for group, camera, and clipping flags with revision-checked writes and rollback.
• Stable IDs (object.id.get / effect.id.get): SDK int64 identifiers.
• MCP Phase 1: the aviutl2-mcp entry point serves nine hub-spoke tools over stdio (pip install 'aviutl2-api[mcp]'). Host-gated, fail-closed, and bundled with this repository.

The 1.0 release gate now only waits for sdk_undo_redo; see protocol/CAPABILITIES_0.10.0.json.

What's New in 0.9.7

0.9.7 adopts the 2026-09-05 official SDK mirror and adds host-gated native editing and rendering. Every new SDK path requires a minimum AviUtl2 host version; older hosts keep the verified 0.9.6 alias fallbacks and fail closed with explicit errors instead of touching out-of-bounds SDK members.

• Native effect reorder (object.effect.reorder via move_effect) on AviUtl2 2.1.3+: the object identity stays stable, verified alias replacement remains the fallback on older hosts.
• Native media trim (media.trim via section endpoint moves) on 2.1.4+: collision preflight, LIFO rollback, read-back verification; covers multi-section media at fixed speed.
• Frame marks API (mark.list/mark.set/mark.clear/mark.move) on 2.1.4+: revision-checked, occupancy-validated, explicitly non-undoable.
• Object-level rendering (render_object_frame / render_object_audio) on 2.1.3+: revision-bound PNG and stereo f32le PCM for a single object.
• Effect item types 17-19 (number_group, group, separator) are now reported by inspection.
• LiveClient exposes host capability flags (sdk_frame_marks, sdk_move_effect, sdk_object_rendering, sdk_section_endpoints) so agents can feature-detect at runtime.

See the compact Agent API Card, the Agent Quick Start, the complete API manual, and the v0.9.7 release notes.

Safe Python Quick Start

Use the same root imports for local files, an open AviUtl2 window, and explicit synchronization. None of these operations saves an .aup2 implicitly.

from aviutl2_api import EditPlan, LiveProject, LocalProject, SyncSession, effect

local = LocalProject.load("project.aup2")
plan = EditPlan().add_text(
    "第一章",
    key="title",
    duration=90,
    y=-200,
    effects=[effect("glow", strength=50)],
)

with LiveProject.connect(pid=46016) as live:
    sync = SyncSession.bind(local, live)
    result = sync.apply(plan)  # explicit Live + local-memory mutation
    native_png = live.render(result.objects["title"].primary.midpoint).png

checkpoint = local.checkpoint()  # project.ai-0001.aup2; source unchanged

For local-only code, immediate methods are shorter and have the same EditPlan semantics:

local = LocalProject.load("project.aup2")
title = local.add_text("第一章", duration=90, y=-200)
title = local.update(title.primary, x=120, scale=110)
local.checkpoint()

Replacing the loaded source is intentionally verbose and must be explicitly authorized: local.save_source(overwrite=True, backup=True).

Installation

pip install aviutl2-api

Advanced direct-file CLI

The CLI predates LocalProject and directly writes the path supplied to its editing commands. Agents should prefer the guarded Python workflow above unless the user explicitly selected direct CLI mutation.

# Create new project
aviutl2 new project.aup2 --width 1920 --height 1080 --fps 30

# Add objects (frame range is optional - defaults to 60 frames, auto-appends)
aviutl2 add text project.aup2 "Hello World"                    # Auto: frames 0-59
aviutl2 add shape project.aup2 circle --duration 90            # Auto: frames 60-149
aviutl2 add text project.aup2 "Manual" --from 0 --to 90        # Manual: frames 0-90

# View timeline
aviutl2 timeline project.aup2

# Apply preset
aviutl2 preset init                         # Initialize sample presets
aviutl2 preset apply project.aup2 0 fade-in # Apply preset to object

# Preview frame (for Vision AI)
aviutl2 preview project.aup2 --frame 0 -o preview.png
aviutl2 preview project.aup2 --frame 0 -o small.png --max-width 800  # Resized for API

# Add animation
aviutl2 animate project.aup2 0 opacity --start 0 --end 100 --motion smooth

# Add filter
aviutl2 filter add project.aup2 0 blur --strength 10

# Batch edit (regex filtering)
aviutl2 batch project.aup2 --filter-text "Hello.*" --color ff0000  # Change all "Hello" texts to red
aviutl2 batch project.aup2 --filter-layer "1-5" --opacity 50       # Set opacity for layers 1-5

# Fix collisions (auto-resolve layer conflicts)
aviutl2 fix project.aup2                                           # Detect and auto-fix collisions
aviutl2 fix project.aup2 --dry-run                                 # Check only (no changes)

Advanced mutable model API

The parser/model/serializer interface is retained for compatibility and custom format work. serialize_to_file() writes the requested destination directly; it does not provide LocalProject checkpoint and hash guards.

from aviutl2_api import parse_file, serialize_to_file, to_json

# Load project
project = parse_file("my_project.aup2")

# Access scenes and objects
scene = project.scenes[0]
for obj in scene.objects:
    print(f"Layer {obj.layer}: frames {obj.frame_start}-{obj.frame_end}")

# Save project
serialize_to_file(project, "output.aup2")

# Export as JSON
json_data = to_json(project)

Live Bridge (0.10.0 beta)

LiveProject is the standard entry point for short, revision-safe edits of the project currently open in AviUtl2:

from aviutl2_api import LiveProject, effect

with LiveProject.connect(pid=12345) as project:
    print(project.describe_schema("glow"))
    title = project.add_text(
        "第一章",
        duration=90,
        y=-200,
        size=64,
        effects=[
            effect("glow", strength=50, color="#FFD966"),
            effect("outline", size_px=4, color="#202040"),
        ],
    )
    title = project.update(
        title.primary,
        x=120,
        scale=110,
    )
    preview = project.render_preview(title.primary, max_width=480)
    image_bytes = preview.data
    image_mime = preview.mime_type

For photos, the high-level API can calculate a scene-relative scale and apply the common non-mirrored EXIF orientations without rewriting the source file:

photo = project.add_image(
    "photo.jpg",
    fit="contain",
    apply_exif_orientation=True,
)

project.get_snapshot() always reads fresh state; project.snapshot is the cached property and is not callable. Use find_objects() for partial file/text matching and frame-range overlap filters. Explicit placement conflicts expose the conflicting object IDs and a suggested free layer.

Omitted at uses the GUI cursor frame; omitted layer selects the first unlocked, collision-free layer from Layer 0. Multiple edits can be validated and applied as one GUI Undo unit with EditPlan:

from aviutl2_api import EditPlan, LiveProject, effect, linear

plan = EditPlan(sequence="parallel")
plan.add_video("intro.mp4", key="intro")
plan.add_text(
    "第一章",
    key="title",
    duration=90,
    y=-200,
    size=80,
    effects=[effect("glow", strength=50)],
)
plan.add_shape(
    "star",
    key="star",
    duration=90,
    x=linear(900, -900),
    rotation=linear(0, 360),
)

with LiveProject.connect(pid=12345) as project:
    validation = project.validate(plan)
    if not validation.valid:
        raise RuntimeError(validation.errors)
    result = project.apply(plan)
    title = result.objects["title"].primary

The bridge delegates canonical Alias parsing, media loading, edits, and renders to AviUtl2's official SDK. It is local Windows-only and does not interpret natural language. LiveProject.client is the documented escape hatch for full Alias, raw item values, manual revisions, and other advanced endpoints. AI agents should start with docs/LIVE_BRIDGE_AGENT_QUICK_START.md and open docs/LIVE_BRIDGE_AGENT_API_MANUAL.md only when a complete reference or advanced operation is needed.

Generating or saving a PNG does not prove that an AI inspected it. Before an agent reports completion, attach RenderedPreview.data with its mime_type to the framework's actual vision input, review several frames or a contact sheet, apply corrections, and render again.

External API access starts disabled in every AviUtl2 process. Open that window's 設定 > 外部API連携設定... menu, then check このウィンドウの外部API連携を許可 in the independent settings window. It includes the target PID and reports either:

• Disabled（接続を受け付けません）
• Enabled（接続受付中）
• Enabled（起動失敗・ログを確認）

The window is hidden at startup and is not docked into AviUtl2's editing workspace. Closing it only hides settings; it does not implicitly change the current permission.

Permission lasts only for that AviUtl2 process. Disabling it stops that process's named pipe and removes its discovery entry. If more than one process is enabled, automatic selection is rejected and the target must be explicit:

from aviutl2_api.live import LiveClient

with LiveClient.connect(pid=12345) as client:
    print(client.get_project_info())

Advanced: low-level LiveClient

Use the following APIs when full Alias data, raw localized item values, explicit revision handling, or an endpoint not wrapped by LiveProject is required.

Existing objects can be edited through revision-scoped snapshots:

with LiveClient.connect(pid=12345) as client:
    snapshot = client.get_snapshot()
    title = snapshot.objects[0]

    client.set_text(title, "更新したタイトル")

    # Every successful edit creates a new revision.
    title = client.get_snapshot().objects[0]
    client.set_position(title, x=120, y=-45)  # X/Y are one Undo unit

    title = client.get_snapshot().objects[0]
    client.move_object(title, layer=2, frame=90)

    title = client.get_snapshot().objects[0]
    client.delete_object(title)

If a user changes the project between snapshot and mutation, the bridge returns STALE_PROJECT_STATE instead of guessing the target.

To protect an object from AI/API edits, select it in the AviUtl2 timeline and use 外部API編集ロック > ロック in the object context menu. The timeline label is prefixed with 🔒 while retaining the current text/effect label, and snapshot.objects[n].api_locked becomes True. set_item(s), move, and delete then fail with OBJECT_API_LOCKED. This lock does not block normal GUI editing; use 外部API編集ロック > 解除 to remove it. Locking or unlocking changes the project revision, so capture a fresh snapshot before the next API operation.

Media files can be checked and placed through AviUtl2's own input plugins, and objects can be inspected without parsing Japanese Alias keys by guesswork:

with LiveClient.connect(pid=12345) as client:
    probe = client.probe_media(r"C:\media\clip.mp4")
    if probe.readable:
        created = client.add_video(
            r"C:\media\clip.mp4",
            layer=2,
            frame=120,
            length=0,  # Let AviUtl2 choose the native duration.
        )
        print(created.frame_start, created.frame_end)

    snapshot = client.get_snapshot()
    details = client.inspect_object(snapshot.objects[0])
    for effect in details.effects:
        print(effect.selector, effect.enabled, effect.locked)
        for item in effect.items:
            print(item.name, item.type, item.raw_value, item.track)

    client.set_property(
        snapshot.objects[0],
        effect="標準描画",
        item="X",
        value=240.0,
    )

Plugin/client 0.7 adds host-discovered effects and layers plus practical clip editing:

with LiveClient.connect(pid=12345) as client:
    catalog = client.get_effect_catalog()
    layers = client.get_layers()
    print(catalog.effects[0].name, layers.layers[0].locked)

    clip = client.get_snapshot().objects[0]
    client.set_playback_rate(clip, 2.0)  # AviUtl2 native 200%; length unchanged

    clip = client.get_snapshot().objects[0]
    split = client.split_media(clip, frame=90)
    print(split.left, split.right)

split_media() is intentionally conservative: it supports basic video/audio clips with fixed playback position/speed and refuses animated or multi-section clips instead of guessing. Effect creation and deletion are available through add_effect() and delete_effect(); deletion uses the selector returned by inspect_object(). duplicate_object() clones an Alias to an empty destination and verifies the resulting snapshot.

0.9.2 adds a practical agent workflow without exposing project save/export or playback controls:

from aviutl2_api.live import (
    EditingSession,
    LiveClient,
    SubtitleLayerPolicy,
)

with LiveClient.connect(pid=12345) as client:
    editing = EditingSession(client)
    report = editing.preflight(audio_range=(0, 299))
    if not report.ready:
        print(report.errors)

    client.add_subtitles(
        "captions.vtt",
        layer_policy=SubtitleLayerPolicy(
            base_layer=10,
            max_layers=3,
        ),
        language="ja",
    )

    snapshot = editing.refresh()
    sheet = client.render_review_contact_sheet(snapshot=snapshot)
    audio = client.render_audio(
        frame_start=0,
        frame_end=299,
        expected_revision=snapshot.revision,
    )
    print(audio.analyze(), len(sheet.png))

Frame/contact-sheet/PCM results stay in memory by default. Saving one refuses an existing destination unless overwrite=True. Eight clients can observe the same enabled AviUtl2 process, but SDK work is FIFO-serialized and mutations are session-idempotent.

0.9.3 fixes live snapshots for media and other objects that occupy multiple timeline layers. Such objects are returned once on their base layer instead of causing INVALID_HOST_OBJECT_RANGE.

0.9.4 adds LiveProject, backend-neutral EditPlan, automatic cursor/layer placement, compact object search, and mixed edit.plan.validate/apply. Native mixed plans validate their final layout first, run inside one edit section, and return an explicit best-effort rollback receipt on failure. They intentionally report atomic=False because the current SDK cannot guarantee rollback of every linked object created by third-party media/effect implementations.

0.9.5 adds a shared semantic Effect API for Live Bridge and .aup2 models. Twenty curated profiles use natural units and are checked against the running host's exact catalog before editing:

from aviutl2_api.editing import effect

title = project.add_text(
    "第一章",
    effects=[
        effect("glow", strength=50, color="#FFD966"),
        effect("outline", size_px=4, color="#202040"),
    ],
)
blur = project.apply_effect(title.primary, effect("blur", radius_px=8))
title_object = project.find(text="第一章").one()
blur = project.update_effect(
    title_object,
    blur,
    effect("blur", radius_px=12),
)
print(project.available_effect_profiles())

Native media may be returned as separate video/audio objects or as one combined 映像再生 object. effects= routes both forms; a combined MP4 legitimately reports the same object ID for video- and audio-domain Effects.

Unknown or third-party effects remain available explicitly through native_effect() or LiveProject.client; their meaning is never guessed. Create-time stacks preserve order, duplicates and enabled state and remain one GUI Undo unit. A 0.9.4 plugin refuses a create-time stack instead of splitting it into several Undo operations.

The same Effect specification can be applied to an in-memory .aup2 model:

from aviutl2_api import apply_effects, validate_standard_effects
from aviutl2_api.editing import effect

apply_effects(
    project_model,
    timeline_object,
    effect("glow", strength=50, color="#FFD966"),
    effect("outline", size_px=4),
)
assert validate_standard_effects(project_model).valid

apply_effects() does not save a file. It inserts complete versioned templates after standard drawing/playback Effects, matching AviUtl2's Open/Save canonical order, and renumbers Effect IDs. The manifest ID remains 2001901; the explicit compatibility allow-list accepts generated 2001901 projects and AviUtl2 Open/Save output upgraded to 2010200. Unknown future project versions still fail closed. Disabled standard Effects use AviUtl2's canonical effect.disable=1 marker. The manual Open/Save gate is available in tests/manual/aup2_effect_roundtrip.py.

The current official SDK cannot delete/duplicate scenes or execute Undo/Redo. Scene listing, creation, switching, project create/open/save, and file export are native on 2.1.10+. Scene delete/duplicate and undo/redo fail with SDK_METHOD_UNAVAILABLE, and the 1.0 release gate still waits for sdk_undo_redo. See protocol/CAPABILITIES_0.10.0.json.

Render an exact composite frame with the running AviUtl2 process:

with LiveClient.connect(pid=12345) as client:
    rendered = client.render_frame(
        60,
        output_path="frame-0060.png",
    )
    print(
        rendered.width,
        rendered.height,
        rendered.revision,
        rendered.sha256,
    )

This uses AviUtl2's native scene renderer. Pillow/OpenCV are not used to reconstruct the scene; PNG compression and verified chunk transport happen after AviUtl2 returns its RGBA output.

For the smallest runnable high-level example, enable exactly one Live Bridge window and run:

python examples/live_add_title.py

It adds one title at the GUI cursor, automatically selects an unlocked free layer, and renders the title's midpoint with AviUtl2's native renderer. The PNG stays in memory and only its dimensions and SHA-256 are printed. PID selection, snapshot refresh, revision checks, operation IDs, and the one-command EditPlan are handled by LiveProject.

The shooting-star example shows high-level animated shapes and text:

python examples/live_shooting_star.py

It uses only LiveProject, EditPlan, add_shape(), add_text(), and the localized-name-free linear(start, end) transform. Cursor placement, free-layer selection, Alias generation, revision checks, and one grouped Undo operation are automatic. The native three-frame contact sheet stays in memory unless the caller explicitly invokes sheet.save(...).

MCP Server (stdio)

aviutl2-api ships an MCP Phase 1 stdio server (aviutl2-mcp) so MCP clients can drive the Live Bridge without writing Python:

pip install 'aviutl2-api[mcp]'
aviutl2-mcp

Register it in the MCP client (for example Claude Desktop) as a stdio server running <python> -m aviutl2_api.mcp.server. The server is Windows-local: it attaches to ONE running AviUtl2 process with External API permission enabled (see above), serves nine hub-spoke tools (bridge_card, bridge_status, bridge_call, bridge_help, bridge_find, bridge_snapshot, bridge_render_object, bridge_render_object_audio, bridge_watch_events), derives method exposure from the running host capability manifest, and fails closed before any pipe traffic. Full setup, tool catalog, and examples: docs/LIVE_BRIDGE_AGENT_API_MANUAL.md.

CLI Commands

Project Operations

Command	Description
new	Create new project
info	Show project information
timeline	Display ASCII timeline
preview	Render frame to PNG for Vision AI
layers	List layers
objects	List objects
search	Search objects at frame
range	List objects in frame range
check	Check if placement is possible

Object Operations

Command	Description
add text	Add text object
add shape	Add shape object
add audio	Add audio file
add video	Add video file
add image	Add image file
move	Move object position
delete	Delete object
copy	Duplicate object
modify	Change object properties
batch	Batch edit with filters (regex)
fix	Auto-fix layer collisions

Animation & Effects

Command	Description
animate	Set animation on property
filter add	Add filter effect

Preset System

Command	Description
preset list	List available presets
preset show	Show preset details
preset apply	Apply preset to object
preset save	Save object settings as preset
preset delete	Delete preset
preset init	Initialize with sample presets

JSON Conversion

Command	Description
export-json	Export project to JSON
import-json	Import project from JSON

Frame Preview (Vision AI Integration)

Render project frames to PNG images for verification by Vision-enabled LLMs.

# Render single frame
aviutl2 preview project.aup2 --frame 0 -o preview.png

# Resize for Vision AI (recommended to avoid API size limits)
aviutl2 preview project.aup2 --frame 0 -o small.png --max-width 800
aviutl2 preview project.aup2 --frame 0 -o small.png --max-height 600
aviutl2 preview project.aup2 --frame 0 -o half.png --scale 0.5

# Render filmstrip (multiple frames in one image)
aviutl2 preview project.aup2 --strip --interval 30 -o timeline.png

Resize Options

Option	Description
--max-width N	Limit width to N pixels (maintains aspect ratio)
--max-height N	Limit height to N pixels (maintains aspect ratio)
--scale X	Scale factor (e.g., 0.5 for 50% size)

Warnings: The tool automatically warns when:

• Aspect ratio would be changed
• Scale factor is below 50% (text/lines may become hard to read)
• Scale factor is below 25% (details may be lost)

Sample Presets

17 sample presets are included:

Animations:

• spin-fade-out - Rotate 10 times while fading out
• fade-in, fade-out - Opacity transitions
• slide-in-left, slide-in-right, slide-out-right - Slide animations
• bounce-vertical, bounce-horizontal - Bounce effects
• zoom-in, zoom-out - Scale animations
• spin-once - Single rotation
• orbit - Circular motion

Effects:

• shake - Vibration effect
• glow-pulse - Glow effect
• blur-soft - Soft blur
• text-shadow - Drop shadow for text
• border-white - White border

Development

Setup

# Clone and setup
git clone https://github.com/Marble-GP/AviUtl2_API.git
cd AviUtl2_API
python -m venv .venv

# Activate virtual environment
# Linux/macOS/WSL:
source .venv/bin/activate

# Windows PowerShell:
.\.venv\Scripts\Activate.ps1

# Windows Command Prompt:
.\.venv\Scripts\activate.bat

# Install in editable mode
pip install -e ".[dev]"

Important: Always activate the virtual environment before running aviutl2 commands. Your prompt should show (.venv) when activated.

If you get a PowerShell execution policy error on Windows:

Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser

Testing and Linting

# Test
pytest

# Type check
mypy src/

# Lint
ruff check src/

Publishing to PyPI

.github/workflows/publish.yml builds and validates the Python distributions and the Windows .aux2 plugin on GitHub-hosted runners for pull requests and pushes to main. Pushing a version tag publishes the verified Python artifact to PyPI using a project-scoped API token and creates a GitHub Release containing AviUtl2LiveBridge.aux2 and its SHA-256 checksum:

git tag v0.10.0
git push origin v0.10.0

The tag without its leading v must exactly match project.version in pyproject.toml. Store a project-scoped PyPI API token as the PYPI_API_TOKEN secret in the GitHub pypi environment. The publish action uses PyPI project aviutl2-api; its API-token username is the standard __token__, so no account username or alias is stored in the workflow.

An existing version can be recovered without moving its tag by running the workflow manually and entering the exact publish_version. Leaving the input empty performs validation only and never publishes.

The plugin is compiled and tested entirely on the Actions windows-2022 runner, which matches the Visual Studio 2022 CMake preset; a local Visual Studio installation is not required for publishing. If docs/releases/<tag>.md exists, the workflow uses it as the GitHub Release body; otherwise it falls back to GitHub-generated notes.

Documentation

Start here:

• Live Bridge Agent Quick Start - Default LiveProject/EditPlan workflow for AI agents
• Live Bridge Agent API Manual - Complete Python API, safety rules, errors, and advanced operations
• v0.10.0 Release Notes -
• v0.9.7 Release Notes -
• v0.9.6 Release Notes - Upgrade steps, high-level API examples, compatibility, and known constraints

Contributor references:

• Live Bridge Protocol - Low-level wire contract for client/plugin implementers
• Live Bridge Development - Native build, tests, security regression, and manual integration
• .aup2 Format Notes - Observed file format and parser/serializer compatibility boundary

CLI syntax is available from aviutl2 --help and each subcommand's --help.

License

MIT