AI Content Factory: 10 Open-Source GitHub Repos, Checked
A viral post lists 10 open-source repos for an AI content factory. We checked each: code and weights licenses, activity, the paid APIs several depend on, and which ones download from social platforms.
On October 17, 2026 @He1s_Sammy posted a list of ten open-source repositories for "an entire AI content factory": research, scripts, images, voiceovers, music, short videos, podcasts and social posts, all without hiring a team.
We checked all ten on GitHub and on their model cards: who made them, the license of the code and of the model weights, how active they are, and what each one needs before it does anything. The short version: the code is open source in every case, but "open source" covers very different things here. Three are models you run on your own hardware, and the license of the weights matters more than the license of the code. Several others are front ends for paid APIs. And three of them download or collect material from social platforms, which is a question of those platforms' terms, not of the repo's license.
The original post by @He1s_Sammy on X (October 17, 2026)
The links in the post are shortened on X. These are the full repositories: OpenCreator, Social Media Research Skills, Local Deep Research, FLUX.2, Kokoro, ACE-Step 1.5, Short Video Factory, AI YouTube Shorts Generator, Podcastfy and Social Media Agent.
Code license versus weights license
The post gets the headline license right: FLUX.2's README marks the [klein] 4B models as Apache 2.0 and says Klein 4B fits in about 8GB of VRAM. What it leaves out is that the rest of the family is not: the [klein] 9B models and the 32B FLUX.2 [dev] use the FLUX.2-dev Non-Commercial License. If you pick a FLUX.2 model for thumbnails you sell or post for a client, pick a 4B one or buy a license. Hosted FLUX models are listed with their prices in our FLUX API entry.
Kokoro's model card says it has 82 million parameters and Apache-licensed weights, which matches our Kokoro entry: free, self-hosted, commercial use allowed. ACE-Step 1.5's weights are MIT on Hugging Face, and its model card says, as reported by the ACE-Step team, that the generated music can be used commercially, that it was trained on licensed and royalty-free music, and that it runs locally with less than 4GB of VRAM.
The ten, checked
| Repo | License | Last push | Last release | Needs |
|---|---|---|---|---|
| OpenCreator | Apache-2.0 | Oct 18, 2026 | v3.2.8, Oct 18, 2026 | Codex with a ChatGPT sign-in or API key |
| Social Media Research Skills | MIT | Aug 26, 2026 | none published | A ScrapeCreators API key |
| Local Deep Research | MIT | Oct 18, 2026 | v1.10.7, Aug 28, 2026 | A local or cloud model and a search engine |
| FLUX.2 | Apache-2.0 code; weights per model | Mar 12, 2026 | none published | A GPU; 4B for commercial use |
| Kokoro | Apache-2.0 code and weights | Aug 6, 2025 | none on GitHub; model v1.0 Jan 27, 2025 | CPU or GPU |
| ACE-Step 1.5 | MIT code and weights | Oct 11, 2026 | v0.1.8, May 18, 2026 | A GPU (Mac, AMD, Intel or CUDA) |
| Short Video Factory | AGPL-3.0 | Oct 16, 2026 | v1.3.0, Oct 16, 2026 | Desktop app; README in Chinese |
| AI YouTube Shorts Generator | MIT | Oct 11, 2026 | none published | MuAPI key, or OpenAI or Gemini in local mode |
| Podcastfy | Apache-2.0 | May 4, 2026 | v0.4.0, Nov 16, 2024 | LLM and text-to-speech keys, or local models |
| Social Media Agent | MIT | Oct 15, 2026 | none published | Anthropic, LangSmith, FireCrawl and Arcade keys |
Licenses, push and release dates from GitHub's API, read October 18, 2026; weights licenses from the model cards.
Three things stand out in that table. Kokoro's repository has not had a push since August 6, 2025, more than a year ago, although the model still gains stars every month. Podcastfy's last release is from November 2024. And Local Deep Research has 1,204 open issues on GitHub, a large backlog for a project with one main author.
Stars and momentum
The fastest riser is OpenCreator, formerly KrillinAI: it gained 1,256 stars in September and 1,179 so far in October, against 188 in April. Social Media Research Skills is the newest, created in June 2026, and reached 4,093 stars in about four months, but it has two contributors and its last push was on August 26.
Scraping and downloading: check the platform's terms
Three tools on the list take material from social platforms:
- Social Media Research Skills collects public posts, comments, transcripts and ads from TikTok, Instagram, YouTube, Reddit, X, LinkedIn and others through the ScrapeCreators API. Its README says ScrapeCreators extracts public social data only.
- OpenCreator has a video downloader for YouTube, TikTok, Instagram, Facebook, Reddit and other sites, and its README notes that some sources may require platform cookies.
- AI YouTube Shorts Generator fetches the source video from YouTube, through MuAPI by default or with yt-dlp in local mode.
We did not review those platforms' terms for this guide. Whether you may download or collect their content, and whether you may republish a clip of someone else's video, depends on each platform's terms and on who owns the content. That is your responsibility, not the tool's: the open-source license of a downloader says nothing about the rights to what it downloads. Clipping your own long videos into Shorts is the safe case.
What the post gets right
- The licenses it states are correct. Klein 4B weights are Apache-2.0, Kokoro is open-weight and small, and ACE-Step 1.5 runs locally.
- The descriptions match the READMEs. OpenCreator does use Codex CLI as its engine, and the research skills do work with Claude Code and Codex.
- Every repo is real and maintained by its named owner. None of the links points to a fork or a mirror.
What it leaves out
- "Without hiring a team" still means paying for APIs. OpenCreator needs a ChatGPT sign-in or an API key for real tasks. The research skills need a ScrapeCreators key. The Shorts generator uses a MuAPI key by default, or OpenAI or Gemini in local mode. Social Media Agent needs keys for Anthropic, LangSmith, FireCrawl and Arcade even in its basic setup. Podcastfy calls cloud text-to-speech such as OpenAI, Google or ElevenLabs unless you set up local models.
- Weights licenses differ inside one family (FLUX.2), as shown above.
- Copyleft. Short Video Factory is AGPL-3.0.
- Activity. Kokoro's repo is quiet and Podcastfy has not released since 2024.
- Posting is human-in-the-loop by design. Social Media Agent asks you to approve or edit each post and handles platform sign-in through Arcade; it is not a fully automatic publisher.
AgentGid data
- Voiceover: Kokoro is one of the open models in our voiceover catalog, next to Piper (GPL-3.0 engine), Chatterbox (MIT), F5-TTS and Coqui XTTS (weights non-commercial) and VoiceStudio (AGPL-3.0 app). Our VoiceStudio guide compares a local voice studio with ElevenLabs.
- Libraries in agent-built projects: among 6,152 public GitHub projects set up for AI coding agents that we scan, yt-dlp is a dependency in 15 and edge-tts in 16; kokoro-onnx appears in 3, the kokoro Python package in 2 and Kokoro.js in 1.
- Music: ACE-Step 1.5 is not yet in our AI music catalog; the only open-source entry there today is Meta's AudioCraft, whose weights are non-commercial.
- Scraping for the agent: Social Media Agent reads pages through Firecrawl, which we track as an agent tool.
If you build it with a coding agent
The research skills install as Agent Skills for Claude Code, Codex, Cursor, Gemini CLI and others, so a coding agent can call them inside your own pipeline. OpenCreator goes the other way: it wraps Codex in a desktop workspace with creator tools.
For a pipeline you own end to end, AgentGid's Short-video kit turns a script into a vertical video with voiceover, stock footage and captions, and keeps a record of every license, which is the part most of these repos leave to you. For editing rather than generating, see AI video editing: 10 open-source repos, checked.
Where to start
- Voiceover on your own machine: Kokoro, with Apache-2.0 weights; accept that the repo is quiet.
- Music for videos: ACE-Step 1.5, MIT weights, after reading its model card on commercial use.
- Thumbnails: FLUX.2 [klein] 4B, the Apache-2.0 member of the family.
- Research: Local Deep Research if you want it local; budget time for its issue backlog.
- A workspace that does many jobs: OpenCreator, if you already pay for ChatGPT or an API key.
- Social research and clipping: read the platforms' terms first, then try the research skills or the Shorts generator on your own content.
We did not run these tools; everything above comes from the repositories, their README and license files, the Hugging Face model cards and GitHub's API, read October 18, 2026, and from AgentGid's own catalog.