The open-source AI workspace & Skills for creators
Bring visual creator tools, reusable Skills, and Agents together for scripts, video, images, voice, avatars, translation, and editing—all in one workspace.
OpenCreator is built for individuals and teams who want to keep creative and development work running locally. Instead of reimplementing an Agent loop, it uses Codex CLI as the execution engine and adds a stable local Runtime, a visual workspace, and a Desktop host around it.
The product brings together two connected workflows:
AI content creation: use dedicated creator tools for video translation and downloading, image and video generation, voiceovers, article and social post writing, short-video scripting, and stick figure animation.
General Agent workspace: organize conversations by project, keep Runs working in the background, and manage approvals, attachments, files, Skills, MCP, schedules, notifications, memory, and diagnostics from one place.
Web is the single frontend implementation. Desktop loads the same Web build and adds only capabilities that require the operating system, such as directory selection, window lifecycle, tray behavior, and native notifications. With the same data and content viewport, both platforms share the same general UI and Runtime behavior.
Project Highlights
🤖 Codex Native: Reuse the Codex Agent loop, models, reasoning, tool calls, conversations, Skills, and MCP without maintaining a second execution engine.
🚀 Ready-to-Use Desktop App: Launch OpenCreator directly from the desktop app with Codex CLI included; the local Runtime starts on demand and prepares a default project automatically.
🔄 Managed Runtime Components: Inspect bundled, active, and latest yt-dlp versions, check for updates periodically, and update manually while keeping the current working version available if an update fails.
🎨 Multimodal Creation: Create and manage video, images, audio, subtitles, and documents through one connected workflow.
🔗 Dual-Mode Workflow: Work through either the visual workspace or Agent conversation while one shared state machine keeps steps, progress, and results synchronized.
🕘 Versioning: Every revision creates a new version while preserving earlier settings and outputs for review and comparison.
🧩 Reusable Skills: Use video workflow Skills, extend the Agent with your own Skills, and manage MCP through Codex-native configuration.
🧠 Memory: Keep global, project, and thread memory with summaries and reproducible Run input snapshots.
🔐 Local Security: Keep data, attachments, and logs local by default, with approvals and redacted diagnostics.
🌐 Localized Interface: Use the Web or Desktop client in Simplified Chinese, English, or Swedish, with automatic system-language detection or manual selection.
Creator Tools
The current release includes ten creator tools. Available models and services depend on your local Codex environment and AI service settings.
Open the Dashboard to write articles, Xiaohongshu posts, or short-video scripts; create stick figure animations; translate or download videos; generate thumbnails, images, or videos; and create voiceovers with Smart Dubbing.
More creator tools are continuously being added.
Workspace
Status
Capabilities
Video Translation
✅ Available
Import local or public videos; transcribe with cloud or local Whisper services; use LLM context for subtitle segmentation, alignment, terminology, and translation; configure bilingual subtitles, dubbing or a custom voice sample, subtitle styles, landscape or portrait composition, and export SRT, audio, or video
Video Downloader
✅ Available
Parse YouTube, Bilibili, and other supported public links, inspect available quality and format options, and download video or audio for later workflows
Thumbnail Generator
✅ Available
Combine a topic, video link, and optional reference image to generate and compare multiple content-thumbnail variations
Image Generation
✅ Available
Generate with GPT Image from a prompt and optional reference image, configure the aspect ratio and output count, then preview and download individual images
Article Writer
✅ Available
Turn a topic, links, videos, or source documents into editable topic options, an outline, and a complete article; add generated images and export Markdown, HTML, or PDF
Xiaohongshu Posts
✅ Available
Generate a complete Xiaohongshu post from a topic or source material, with controls for the target audience, post type, and length, then copy or download the result
Short Video Script
✅ Available
Create a shoot-ready segmented script from a topic or source material, tailored to the audience, platform, duration, and tone, then edit, copy, or download it
Stick Figure Animation
✅ Available
Turn text or YouTube content into narration, voice, consistent-character storyboard visuals, subtitles, and a downloadable stick figure animation
Auto Clips
In development
Analyze long videos, identify highlights, and turn selected moments into reusable short clips
Smart Dubbing
✅ Available
Turn scripts into voiceovers with selectable voices, pacing, and emotion controls
Video Generation
✅ Available
Generate videos with Seedance from prompts and reference images, then preview, regenerate, and download each version
Digital Avatar
In development
Combine scripts, voice, and avatar presentation to produce talking-head videos
Skills
Creator tools provide visual controls; Skills give the Agent reusable instructions and tool workflows. OpenCreator includes video-production Skills in the repository, alongside support for managing local Codex Skills.
The repository's skills/ directory contains reusable instructions for Agents operating the embedded KrillinAI CLI.
Validate a multi-stage output plan in dry-run mode; execute actual work through the individual stage Skills
Extend with Your Own Skills
OpenCreator supports local Codex Skills defined by SKILL.md, so you can add your own methods and workflows rather than relying only on fixed creator tools. Skill availability depends on the active Codex home and installed Skills; video workflow Skills require the CLI and relevant services to be configured. Inclusion in the repository does not mean every Skill is automatically installed or every external service is bundled.
Conversation and workspace, moving together
Describe tasks naturally, then step into visual tools whenever you need precise control.
Fine-grained workspace controls
Adjust subtitles, shots, audio, and generation settings precisely.
Flexible conversational edits
Tell the Agent what to change and refine the result in natural language.
Synchronized state
Conversation and workspace share the current task state, so nothing needs repeating.
Independent versions
Each revision creates a separate version without overwriting earlier results or settings.
Models Supported
Language model availability follows the Codex model catalog or your OpenAI-compatible provider. Image, video, voice, and transcription models use the services configured in Settings → AI Services.
Language models
GPT
DeepSeek
Qwen
Kimi
GLM
Grok
Doubao
ERNIE
Hunyuan
Image
GPT Image
Video
Seedance
Voice and transcription
Whisper
OpenAI TTS
MiniMax
Edge TTS
Aliyun Speech
Examples
Video Translation
The public examples below were produced when OpenCreator was still named KrillinAI. They demonstrate the established subtitle alignment, translation, dubbing, and portrait-video workflow that OpenCreator's Video Translation workspace brings into a wider Agent workflow.
The project generated the subtitle file below from a 46-minute local video in one run, without manual subtitle adjustments. The published result shows complete coverage, no overlapping lines, natural segmentation, and high-quality translation.
These video examples and the subtitle alignment image were produced while OpenCreator still used the KrillinAI name.
Video Generation
Generate an AI video from a text prompt or reference image with Seedance. Configure the model, aspect ratio, resolution, and duration, then preview, regenerate, or download each version from the project workspace.
Video Downloader
Analyze a public video link, compare the available formats, and download video or audio directly to the project.
Stick Figure Animation
OpenCreator developed this original character collection in collaboration with artist Harbor Hsia, creator of Stickman on Behance. The built-in cast keeps character identities consistent throughout the animation workflow.
Turn text or YouTube content into a complete animation through a guided workflow for script review, narration, timing, storyboard visuals, subtitles, rendering, and downloadable video output.
Quick Start
Prerequisites
Node.js 22 or later
pnpm 9.15.0, pinned through the repository's packageManager field
Open http://127.0.0.1:19861/. The development server starts the local daemon on demand and injects a temporary Runtime token through a same-origin proxy, so no connection token needs to be copied manually.
On first launch, the Runtime prepares a default project. The composer is ready as soon as the connection completes. To work on the daemon only:
pnpm daemon:dev
The daemon listens only on a loopback address and prints its connection address and temporary token to stdout once.
Desktop
Desktop and the browser use the same React frontend from apps/web. General project, conversation, task, and settings behavior calls the same Daemon/API. Electron adds only real system paths, window controls, tray behavior, and native notifications.
Development Mode
pnpm desktop:dev
Local Packaging
Command
Output
pnpm desktop:package
A runnable directory for the current platform, intended for local verification
pnpm desktop:dist
An installer for the current platform
pnpm desktop:release
The formal release packaging entry point
pnpm --filter @opencreator/desktop verify:package
Verification for an existing Desktop package
pnpm krillinai:package
Separate KrillinAI Server and CLI archives for the selected platform
Desktop packaging rebuilds Web from the current workspace, records the commit, dirty state, platform, architecture, and Web hash, and compares apps/web/dist with the resources embedded in the application. Packaging fails if they differ. See the Desktop release runbook for signing, notarization, Windows builds, and release requirements.
Core Workflows
Conversations and Runs
Select a project or start a new conversation.
Enter a task and choose the permission level, Profile, model, and reasoning effort.
While a Run is active, queue follow-up tasks or interrupt it and continue immediately.
Use the Timeline to inspect reasoning summaries, tool calls, file changes, approvals, and final results.
Use the task center to track running, completed, failed, and approval-blocked tasks globally.
Skills and MCP
Browse the Skill marketplace, installation history, and locally available Skills in the plugin center.
Select a Skill from the composer with / or the add menu so the next task follows its workflow.
MCP management passes through Codex-native commands and configuration instead of maintaining a second execution engine.
OpenCreator uses the active $CODEX_HOME by default, so confirm the impact before changing global Skills or MCP configuration.
Schedules and Dedicated Task Threads
Every schedule owns a persistent, dedicated OpenCreator conversation.
Automatic triggers, manual runs, and user follow-ups reuse that conversation and run serially with the queue or skip policy.
Deleting a schedule archives its dedicated conversation while preserving existing Runs, results, and underlying Codex history.
Rotating or recovering an underlying Codex thread does not change the OpenCreator task entry or page route.
OpenCreator System Architecture
OpenCreator treats the visual workspace and the Agent conversation as two interfaces to the same creative task, rather than two separate workflows. Each creator workflow is modeled as a state machine: source input, configuration, generation, review, revision, and export become explicit states and events. Workspace actions and conversational commands enter the same state machine, while the current step, configuration, progress, versions, and results are projected back into both interfaces. This keeps the workspace and conversation synchronized without introducing a second source of truth.
Creative work is iterative, so revisions do not overwrite the current result. Each correction or regeneration creates a new version from the existing workflow state, retaining the settings and outputs of earlier versions for review, comparison, and continued refinement.
Provides the Agent loop, sessions, reasoning, tools, Skills, and MCP
Codex CLI · app-server
Media Toolchain
Downloads, transcribes, transforms, generates, and exports creative media
yt-dlp · Whisper · FFmpeg · configured AI services
Local Data
Stores project data, Runs, attachments, outputs, and credentials locally
SQLite · filesystem · system credential storage
Desktop Host
Loads the shared Web build and adds operating-system capabilities
apps/desktop · Electron · Preload Bridge
Core principles:
The workspace and Agent conversation are synchronized projections of one workflow state; both dispatch events to the same state machine instead of maintaining parallel task state.
Revisions create new versions instead of replacing existing results, preserving the context and output of every creative iteration.
The frontend does not launch Codex directly and does not depend on raw Codex JSONL event formats.
The daemon owns process lifecycle, event normalization, persistence, approvals, schedules, and the notification outbox.
Codex remains the execution source of truth for the Agent loop, Skills, and MCP.
Browser Bridge and Desktop Bridge do not implement separate copies of general product logic.
Repository Layout
OpenCreator/
├── apps/
│ ├── web/ # The single React frontend implementation
│ ├── daemon/ # Local Fastify Runtime and Codex adapter
│ ├── desktop/ # Electron Main, Preload, native capabilities, and packaging
│ └── harness/ # Runtime command-line verification tool
├── packages/
│ ├── protocol/ # Runtime contracts shared by Web, Daemon, and Desktop
│ └── skill-market/ # Skill marketplace models and shared logic
├── docs/ # Design docs, API references, runbooks, and test reports
├── scripts/ # Repository-level checks
└── .runtime/ # Local Runtime data, created on first launch
Configuration
AI Service API Keys
Open Settings → AI Services to configure the model, transcription, voice, and image providers used by the current workspaces. Additional service categories may appear in preparation for upcoming creator tools. Each category exposes only the fields required by its selected provider, including the Base URL, API Key, model, proxy, or provider-specific credentials.
Credentials are saved through the local Runtime's system credential storage and should never be committed to the repository. Some local or system-backed providers, such as Edge TTS, do not require an API Key.
Third-party Runtime Components
Open Settings → Third-party Components to inspect the yt-dlp nightly version currently in use, the version bundled with OpenCreator, its source, and the latest available release. OpenCreator checks for updates every seven days but never installs them automatically. Updates require an explicit user action, and the current working version remains available if downloading, verification, or installation fails.
Runtime Environment Variables
Most users do not need environment variables. Use these when you need isolated data, a specific Codex executable, or a custom managed-project directory:
Environment variable
Default
Purpose
OPENCREATOR_DATA_DIR
.runtime
OpenCreator database, Runs, attachments, and managed workspaces
OPENCREATOR_CODEX_BIN
codex
Path to the Codex CLI executable
CODEX_HOME
~/.codex
Source of truth for Codex sessions, configuration, Skills, MCP, and Profiles
OPENCREATOR_DEFAULT_CWD
Current working directory
Default daemon working directory
OPENCREATOR_DEFAULT_PROJECT_ROOT
Runtime default policy
Managed-project root; when set, OpenCreator uses its OpenCreator/ child directory
OPENCREATOR_CODEX_THREAD_ROTATION_RUN_THRESHOLD
50
Terminal Run threshold for rotating the Codex thread behind a long-running schedule; use 0 to disable proactive rotation
For example, isolate both Runtime data and the Codex environment:
Build dependencies and start Electron in development mode
pnpm test
Run workspace unit and integration tests
pnpm typecheck
Run TypeScript checks across the repository
pnpm build
Build every workspace
pnpm e2e
Run Web Playwright E2E tests
pnpm smoke:ci
Run the fake-Codex Runtime smoke test
pnpm perf:check
Check the recorded performance baseline
Before submitting a change, run at least:
pnpm test
pnpm typecheck
pnpm build
Changes to Desktop, Host Bridge, the Runtime proxy, or shared frontend workflows also require Web/Desktop consistency tests, packaged application E2E, and Web build hash verification. Passing Web unit tests alone does not establish Desktop release readiness.
The real Codex smoke test is disabled by default. Enable it explicitly with:
OPENCREATOR_RUN_REAL_CODEX_SMOKE=1 \
pnpm --filter @opencreator/daemon test -- test/smoke/real-codex-smoke.test.ts
The root README.md is the canonical English document. Maintained translations live at docs/<locale>/README.md. Add a language to the switcher only after its full document has been translated and synchronized with the English structure.
Contributing
Describe the problem, use case, and expected behavior in Issues.
Create a focused feature or fix branch from the latest development branch.
Follow the existing architecture: implement general product capabilities once in Web and Daemon, and isolate native differences behind explicit capabilities.
Add appropriate unit, integration, or E2E coverage for behavior changes, and list both completed and skipped verification in the Pull Request.
Never commit .runtime/, local credentials, Codex sessions, build caches, or other user data.
Contributors
Thanks to everyone who has taken part through code, documentation, feedback, issue reports, Skills, designs, and ideas.
Star History
OpenCreator was formerly named KrillinAI. This chart covers the repository's full history across the rename.
