Open Source
Open Source

Skip the agent loop: 12,078 stars for an open AI workspace

krillinai/OpenCreator is an open-source AI creation workbench and Skills collection maintained by the krillinai team (formerly KrillinAI), released under the permissive Apache-2.0 license and usable commercially and self-hosted. As of a 2026-09-22 GitHub snapshot it shows 12,078 stars, 1,222 forks, TypeScript as primary language, created 2024-12-17, last pushed 2026-09-21 and 31 open issues, figures that are a same-day snapshot rather than a lasting level. Its defining design choice is local-first: project data, attachments and logs stay on your machine by default (SQLite and the file system, with Codex sessions and config in CODEX_HOME), the Daemon listens only on 127.0.0.1 with Bearer tokens required for every endpoint except health checks, HTML preview disables scripts and navigation by default, and the desktop package enables ASAR integrity checks and Cookie encryption. The pivotal architectural judgment is that it does not rebuild the Agent loop but uses Codex CLI directly as the execution engine, wrapping it with three layers, a local Runtime, a visual workbench and a desktop host, so the agent loop, sessions, reasoning, tool calls, Skills and MCP all come from Codex; the upside is not maintaining a second engine, inheriting Codex's capabilities and using Codex-native configuration for Skills and MCP, while the cost is a hard dependency on the Codex ecosystem, a capability ceiling set by Codex, and available models depending on your local Codex and AI service settings. The README's prose says ten creation tools while its own table lists twelve rows, ten available and two in development (Auto Clips and Digital Avatar), plus seven built-in video-production Skills (KrillinAI CLI, Subtitle, TTS, Landscape and Portrait Render, Cover, Pipeline Plan). It is worth stating that a Skill in the repository is not auto-installed nor are external services bundled, and that it is not an open-source replacement for JianYing since its core is Agent plus creation tools plus Skill orchestration without a multi-track timeline editor.

Published September 22, 20268 min read
<!-- opencreator-resource | open-source | Skip the agent loop: 12,078 stars for an open AI workspace -->

What it is: an open workspace that keeps creation and development local

OpenCreator is an open-source project by the krillinai team, formerly named KrillinAI; the repository is now krillinai/OpenCreator. It positions itself as an open-source AI workspace and a set of Skills for creators, bringing visual creation tools, reusable Skills, and Agents into one workspace covering scripts, video, images, voice, avatars, translation, and editing. The GitHub snapshot of 2026-09-22 shows 12,078 stars and 1,222 forks. The primary language is TypeScript. The repo was created on 2024-12-17 and last pushed on 2026-09-21, with 31 open issues. Stars and forks change daily, so read these as a same-day snapshot. It also took the third-party honor "#1 Repository of the Day" on Trendshift, a single-day recognition that is not a long-term or authoritative ranking.

The design keeps both creative and development work local. Your project data, attachments, and logs stay on your machine by default rather than going to a vendor cloud. It offers two connected modes. The content workspace uses visual tools and templates for video translation and downloading, image and video generation, voiceovers, article and social writing, short-video scripts, and stick figure animation. The Agent conversation lets you start and guide a task in natural language, organize by project, run background Runs, and manage approvals, attachments, files, Skills, MCP, schedules, notifications, memory, and diagnostics in one place.

The frontend strategy matters: Web is the only frontend, and the desktop app loads the same Web build, adding only OS-level capabilities like directory selection, window lifecycle, tray behavior, and native notifications. Browser and desktop share the same React frontend and Daemon interface; the desktop only adds an Electron host for system features.

Architecture judgment: no rebuilt Agent loop, Codex CLI reused directly

This is the key point. OpenCreator does not rebuild its own Agent loop. It uses Codex CLI directly as the execution engine and wraps it with three layers: a stable local Runtime, a visual workspace, and a desktop host. The Agent loop, sessions, reasoning, tool calls, Skills, and MCP all come from Codex; OpenCreator only packages that engine into a local product.

The benefits are real. First, no second execution engine to maintain. The Agent loop is the hardest and most maintenance-heavy part, so reusing Codex hands that debt to upstream and frees the team for creation tools and Runtime. Second, model and reasoning follow Codex. Whatever Codex supports, OpenCreator can use, without another round of model adaptation. Third, Skills and MCP use Codex-native configuration. The README says MCP management goes through Codex-native commands, not a second engine, and OpenCreator uses the active CODEX_HOME by default.

The costs are clear too. First, heavy dependence on the Codex ecosystem: any change to Codex interface, licensing, or behavior directly hits OpenCreator. That is a single-point dependency. Second, the ceiling is bounded by Codex; anything Codex lacks, OpenCreator cannot easily add. Third, available models depend on your local Codex environment and AI service settings. The README is explicit: available models and services depend on your local Codex environment and AI service settings. OpenCreator bundles no model of its own; capability comes from Codex CLI and external services.

In short: the smart move is recognizing the Agent loop is not worth rebuilding, choosing leverage over reinvention. But it ties OpenCreator's fate to Codex. This is a clear trade-off, not a free lunch.

Ten available tools, plus two in development

The README text says ten creator tools, but its table lists twelve rows: ten Available and two In development, Auto Clips and Digital Avatar. The table below shows them as written, without rounding.

ToolStatusCapabilities
Video TranslationAvailableImport local or public videos; transcribe with cloud or local Whisper; 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; export SRT, audio, or video
Video DownloaderAvailableParse YouTube, Bilibili, and other public links, inspect available quality and format options, and download video or audio
Thumbnail GeneratorAvailableCombine a topic, video link, and optional reference image to generate and compare multiple thumbnail variations
Image GenerationAvailableGenerate with GPT Image from a prompt and optional reference image, configure aspect ratio and output count, then preview and download individual images
Article WriterAvailableTurn a topic, links, videos, or source documents into editable topic options, an outline, and a complete article; add generated images; export Markdown, HTML, or PDF
Xiaohongshu PostsAvailableGenerate a complete Xiaohongshu post from a topic or source material, with controls for target audience, post type, and length
Short Video ScriptAvailableCreate a shoot-ready segmented script from a topic or source material, tailored to audience, platform, duration, and tone
Stick Figure AnimationAvailableTurn text or YouTube content into narration, voice, consistent-character storyboard visuals, subtitles, and a downloadable stick figure animation
Auto ClipsIn developmentAnalyze long videos, identify highlights, and turn selected moments into reusable short clips
Smart DubbingAvailableTurn scripts into voiceovers with selectable voices, pacing, and emotion controls
Video GenerationAvailableGenerate videos with Seedance from prompts and reference images, then preview, regenerate, and download each version
Digital AvatarIn developmentCombine scripts, voice, and avatar presentation to produce talking-head videos

A few notes. Video translation is the original strength from the KrillinAI era: a forty-six-minute local video yields complete subtitles in one run, with full coverage, no overlap, natural segmentation, and solid translation. Video download uses the managed yt-dlp component; OpenCreator checks updates every seven days but never installs automatically, and keeps the working version if an update fails. Image and video generation go through external services (GPT Image, Seedance) configured under Settings to AI Services. Article writing, Xiaohongshu posts, and short-video scripts are pure-text and export directly. Stick figure animation uses an original cast for consistent characters. Auto Clips and Digital Avatar are In development and should not be treated as stable.

Seven Skills: reusable workflows for the Agent

Creation tools give visual controls; Skills give the Agent reusable instructions and workflows. OpenCreator ships seven video-production Skills in the skills directory, each a folder with SKILL.md.

SkillCapabilities
KrillinAI CLIChoose commands, check configuration, and interpret progress, manifests, outputs, and errors
SubtitleDownload platform captions or transcribe media, translate subtitles, and produce bilingual or short portrait subtitles
TTSGenerate target-language dubbing from subtitles and optionally produce a dubbed video
Landscape RenderRender landscape videos with bilingual subtitles or dubbed audio and target-language subtitles
Portrait RenderCompose portrait videos with titles, bilingual subtitles, or dubbing
CoverGenerate a cover image from a complete text prompt and save the image and final prompt
Pipeline PlanValidate a multi-stage output plan in dry-run mode; execute actual work through the individual stage Skills

It also supports local Codex Skills defined by SKILL.md, so you can add your own methods. But the README warns: including a Skill in the repo does not mean it is auto-installed, nor that every external service is bundled. Video workflow Skills need the CLI and services configured first. Seeing a Skill means it is usable; configuration still stands between that and working out of the box.

Why local-first is valuable

Local-first decides data ownership, compliance, and control. OpenCreator keeps data, attachments, and logs local by default in SQLite and the filesystem, while Codex sessions stay in CODEX_HOME and need separate backup.

Concrete points: approvals and redacted diagnostics, where sensitive memory needs a second confirmation and logs are redacted before export; three-level memory (global, project, thread) with summaries and reproducible Run snapshots; MCP support through Codex-native configuration; and a localized interface in Simplified Chinese, English, and Swedish with auto or manual selection.

Security boundaries add to this: the Daemon listens only on 127.0.0.1 and requires a Bearer token except for health checks; HTML preview disables scripts, navigation, and popups; desktop packages enable ASAR integrity and cookie encryption while disabling RunAsNode, NODE_OPTIONS, and the Node CLI Inspector. Together these make local-first more than a slogan.

Cold thinking: strength, expectation, and single-point dependency

12,078 stars prove strength and raise expectations, but they do not mean it is production-ready. The 31 open issues show active iteration, and Auto Clips and Digital Avatar are still In development.

Not rebuilding the Agent loop is smart yet creates a single-point dependency on Codex; any upstream change pulls OpenCreator along, and users must maintain their own Codex environment and evolve with it.

It is also not a CapCut open-source replacement. Its core is Agent plus creation tools plus Skill orchestration, with no multi-track timeline editor. The loop it closes is running a creation task to completion with language and visual tools, not frame-by-frame timeline editing. Comparing it with CapCut or OpenCut and asking who replaces whom is the wrong question; the right one is how the loop is built. These are two different solutions, not two answers to one problem.

Per site rule, this article only deep-dives OpenCreator and does not repeat existing pieces, neither the OpenCut breakdown nor the AI video production tools comparison. For a CapCut replacement or a multi-tool comparison, read those. This piece answers only how OpenCreator is built, how it thinks, and whether it is worth following.

The four batch articles interlock; the three below link back to this one. Jianying Hub release is the same-batch hotspot, Five AI video workspaces compared is the same-batch comparison, and Jianying Hub hands-on SOP is the same-batch SOP. Links: /en/posts/jianying-hub-hotspot, /en/posts/ai-video-workspace-comparison-review, and /en/posts/jianying-hub-sop. Deferred older pieces: /en/posts/opencut-resource and /en/posts/ai-video-production-tools-comparison-review.

FAQ

Q1: Is OpenCreator completely free and ready to use out of the box?

A1: The software itself is released under the permissive Apache-2.0 license, which allows commercial use and self-hosting and charges no license fee. But it is not completely free: you must provide your own Codex environment and AI services, model calls may incur cost, and first use requires a valid Codex login. Treating it as free software misses that real layer of cost.

Q2: Does it bundle its own large model?

A2: No. OpenCreator is a workspace plus Skills plus a local Runtime. Model capability comes from Codex CLI and external AI service configuration. The README states plainly that available models and services depend on your local Codex environment and AI service settings, so what models you can use depends on what you configured locally.

Q3: The text says ten tools, but I count twelve. Why?

A3: The README text says the current release includes ten creator tools, but the same document's table actually lists twelve rows, of which ten are marked Available and two (Auto Clips and Digital Avatar) are marked In development. This article presents the table as it is, without rounding them off.

Q4: Are the Skills in the repository installed and usable once cloned?

A4: Not necessarily. The README warns clearly: including a Skill in the repository does not mean that Skill is automatically installed, nor that every external service is bundled. Video workflow Skills require the CLI and the relevant services to be configured before they can actually run. Seeing a Skill only means it is usable; configuration still sits between that and working out of the box.

Q5: Will the star and fork numbers stay accurate?

A5: No. The 12,078 stars, 1,222 forks, and 31 open issues here are all GitHub API snapshots from 2026-09-22. The numbers move every day, so treat them only as a same-day snapshot, not as a permanent reading of the project's real standing.

This article is AI-assisted and human-edited. Last updated: 2026-09-22

FAQ

Is OpenCreator completely free and ready to use out of the box?
The software itself is released under the permissive Apache-2.0 license, which allows commercial use and self-hosting and charges no license fee. But it is not completely free: you must provide your own Codex environment and AI services, model calls may incur cost, and first use requires a valid Codex login. Treating it as free software misses that real layer of cost.
Does it bundle its own large model?
No. OpenCreator is a workspace plus Skills plus a local Runtime. Model capability comes from Codex CLI and external AI service configuration. The README states plainly that available models and services depend on your local Codex environment and AI service settings, so what models you can use depends on what you configured locally.
The text says ten tools, but I count twelve. Why?
The README text says the current release includes ten creator tools, but the same document's table actually lists twelve rows, of which ten are marked Available and two (Auto Clips and Digital Avatar) are marked In development. This article presents the table as it is, without rounding them off.
Are the Skills in the repository installed and usable once cloned?
Not necessarily. The README warns clearly: including a Skill in the repository does not mean that Skill is automatically installed, nor that every external service is bundled. Video workflow Skills require the CLI and the relevant services to be configured before they can actually run. Seeing a Skill only means it is usable; configuration still sits between that and working out of the box.
Will the star and fork numbers stay accurate?
No. The 12,078 stars, 1,222 forks, and 31 open issues here are all GitHub API snapshots from 2026-09-22. The numbers move every day, so treat them only as a same-day snapshot, not as a permanent reading of the project's real standing.

Related

Open Source

Univer Office SDK: Agents Edit Cells Without an AI SDK

Only what the dream-num/univer README, LICENSE and GitHub API fields support. Apache-2.0 covers the Web SDK core architecture and engine implementations; commercial plugins, the Server SDK and commercial AI SDK extensions are distributed under separate licences. Facade APIs such as FWorkbook, FWorksheet, FRange, getRange and setValues let an application or Agents Edit Cells directly with formula recalculation, while workbook.save() persists a JSON snapshot rather than XLSX import or export. No AI SDK is required anywhere in that path. Stars are cumulative GitHub API totals, not single-day growth.

Nov 16, 20269 min read
Open Source

An Agent That Lives on Your Phone: Inside the 7,938-Star Mika

wangzan101/Mika is an open-source on-device agent that installs on your Android phone and does the tapping, typing and scrolling itself: a GitHub API snapshot on November 10, 2026 shows 7,938 stars, Kotlin as the primary language, an Apache-2.0 license, and a last push at 04:19 UTC the same day; the November 7, 2026 metadata-review snapshot in our archives counted 7,489 stars, a net gain of 449 with both timestamps attached. The README positioning is blunt: you say what you need, and it works your apps and the web until it is done; the project has a homepage at mikaphone.com and multilingual READMEs. The deep dive covers the background-screen mechanism, Ask and Autonomous modes, three model connection paths, Android 11/12/14 requirements, a built-in app knowledge table, a three-layer privacy boundary, Releases-based installation and a 12-item compatibility self-check, all checked against the README text; unspecified architecture, roadmap and author details are left out rather than invented.

Nov 11, 20269 min read
Open Source

17 Seconds on a CPU: Can an API Prove Which Model Answered?

anthropics/FlashProve is an open-source research prototype for zero-knowledge verifiable inference from Anthropic (Apache-2.0, 843 stars per a November 5, 2026 GitHub API check, version 0.16.0, officially labeled a research artifact: not maintained and not accepting contributions). The three-step protocol: the provider publishes a weight commitment ID offline (about 39 minutes on one H100), attaches a GPU-generated proof to each reply (about 19MB per 1024 tokens), and anyone verifies with cp_verify on a plain CPU in seconds. It runs Llama 4 Maverick at full 400B size on a single H100 at about 110 tokens per second of proof prefill with about 17-second CPU verification, at a cost of roughly 70 to 95 times normal inference GPU time; soundness is formally proved in Lean 4. It proves provenance, not correctness, and no major provider ships proofs yet.

Nov 5, 20269 min read