20 AI GitHub Repos: Licences and Upkeep Checked

OpenClaw, a personal AI agent that runs on your own computer, had about 393,000 stars on GitHub on 28 October 2026, more than any other repository on this list. It is followed by n8n, TensorFlow and AutoGPT. Popularity is the easy part to measure. This list, first published in March and rechecked against each project’s own repository in October, also records what each tool does, which licence it carries and whether it is still being kept up.

Five of the 20 carry licence conditions that matter for commercial use, one is not open source at all, and two are no longer the current version of what they once offered. Star counts below are rounded and dated 28 October 2026, and descriptions of features come from each project’s own README.

Where to start depends on the job

The 20 repositories fall into five rough groups, and few people need all of them. Someone building agents that take actions will spend most of their time in the first group: OpenClaw, AutoGPT, n8n and the agent frameworks. Someone who wants to run models privately on a laptop or a home server needs the second group: Ollama, llama.cpp and Open WebUI. Training or fine-tuning a model means PyTorch, TensorFlow or Transformers. Images, audio and video sit with ComfyUI, Stable Diffusion WebUI and Whisper, and question-answering over a company’s own documents is where RAGFlow and Dify fit.

A sensible order for a newcomer is to run a model locally first, because it shows what a model can and cannot do without a bill attached, then wire it into an app with a framework, and only then add agents that act on their own. Each step adds moving parts, and each step’s licence should be read before it goes into anything commercial.

Agents and automation

OpenClaw (about 393,000 stars, MIT licence) is a personal assistant that runs on your own machine and talks to you through the chat apps you already use. Its README lists Discord, iMessage, Slack, Teams, Telegram, WhatsApp and more than 20 other channels, plus native apps for macOS, iOS, Android, Windows and Linux. A local “Gateway” manages sessions, tools and channel connections, and the models behind it, including Claude, Codex and local models, are plugins that can be swapped. The README says your state and credentials stay local, but prompts still go to whichever model provider you configure, so it is local-first rather than offline. The project is now overseen by the OpenClaw Foundation.

Agents that operate a computer on a user’s behalf are not limited to open source. TSN covered Digital Optimus, the Tesla and xAI agent built on Grok that Elon Musk said would watch a screen and act on it, when it was announced in March.

AutoGPT (about 188,000 stars) began as an early experiment in chaining language-model calls into multi-step tasks. It now presents itself as a “team of AI Experts”, with a coordinating agent called Otto handing work to specialist agents for content, sales research and admin, offered as a hosted trial or self-hosted from a single container. Its platform code is now under the Polyform Shield licence, which allows personal and internal business use but not selling it as a competing hosted service, while the original “classic” code and the rest of the repository remain MIT.

n8n (about 209,000 stars) is a workflow automation tool with more than 1,500 integrations and AI nodes. It combines a visual canvas with JavaScript or Python code, can include human approval steps, runs self-hosted or in n8n’s cloud, and offers more than 9,000 workflow templates. It is “fair-code” under the Sustainable Use License, with enterprise features under a separate licence.

LangChain (about 149,000 stars, MIT) now calls itself “the agent engineering platform”, a framework for wiring models to tools, memory and data through a standard interface, so one model can be swapped for another. Newcomers are pointed to Deep Agents, a higher-level package for planning and sub-agents, and to LangGraph for lower-level control. A JavaScript version, LangChain.js, is maintained separately.

Langflow (about 156,000 stars, MIT) is a visual builder for the same kind of pipelines, with an interactive playground for testing, Python access to every component and the option to expose each flow as an API or MCP server. A desktop version for Windows and macOS bundles its dependencies. CrewAI (about 60,000 stars, MIT) is a Python framework that organises several role-based agents into “Crews” and event-driven automations into “Flows”. A paid suite adds managed deployment and monitoring.

Dify (about 160,000 stars) is a platform for building and running AI apps that combines workflows, a retrieval pipeline, agents, model management and observability in one interface. Its README asks for at least two CPU cores and 4GB of memory to self-host with Docker Compose. Its licence is a modified Apache 2.0 that requires a commercial licence to run it as a multi-tenant service.

Running models on your own hardware

Ollama (about 185,000 stars, MIT) downloads and runs open-weight models with single commands, and shipped an update on 27 October. It installs on macOS, Windows and Linux or runs as a Docker image, has Python and JavaScript libraries, and can now connect coding agents such as Claude Code and Codex, or OpenClaw, to the models it serves.

llama.cpp (about 133,000 stars, MIT) moved to the ggml-org account from its original home at ggerganov/llama.cpp, which still redirects. It is the C/C++ inference engine underneath many local tools, including Hugging Face’s Llama Windows tray app. It has no outside dependencies, treats Apple silicon as a first-class target, supports NVIDIA, AMD and other GPUs, and offers integer quantisation from 1.5 to 8 bits to cut memory use. It can split a model between CPU and GPU when the model is larger than the graphics card’s memory, and its server mode exposes an OpenAI-compatible API. Running models cheaply is the same problem the big chip makers are chasing at data-centre scale. NVIDIA built its GTC 2026 conference around AI inference, as TSN reported in March.

Open WebUI (about 157,000 stars) is a self-hosted chat interface. It can talk to local Ollama models and any OpenAI-compatible API side by side, and adds document search across nine vector databases, user roles and groups, and single sign-on for company deployments. Its Open WebUI License is BSD-based but adds a branding clause for deployments above 50 users in 30 days.

Frameworks and models

PyTorch (about 105,000 stars, BSD-style) and TensorFlow (about 201,000 stars, Apache 2.0) remain the two main deep-learning frameworks. PyTorch’s README describes two core features: tensor computation with strong GPU acceleration, and neural networks built on an automatic differentiation system. TensorFlow was first developed by researchers at Google Brain, and offers stable Python and C++ interfaces plus a smaller CPU-only package. Transformers (about 168,000 stars, Apache 2.0) now describes itself as a model-definition framework for text, vision, audio and multimodal models, so a model defined once works across most training and inference tools.

DeepSeek-V3 (about 105,000 stars, MIT for the code) is the odd one out. It is a mixture-of-experts model with 671 billion parameters in total, of which 37 billion are active for each token, but its repository was last updated on 28 August 2025. For anyone following DeepSeek’s open-source work, the more current release is DeepInfer, the inference engine it open-sourced on 20 October 2026.

Whisper (about 111,000 stars, MIT) is OpenAI’s open speech-recognition model. It transcribes speech in many languages, translates it and identifies the language spoken. It comes in six sizes, from “tiny” at 39 million parameters and about 1GB of video memory to “large” at 1.55 billion parameters and about 10GB.

Images, coding and retrieval

Stable Diffusion WebUI (about 165,000 stars, AGPL-3.0) was the classic local image-generation interface, with text-to-image, image-to-image, inpainting and outpainting in a browser page. But its last code push was on 2 March 2026 and its latest release is still v1.10.1.

ComfyUI (about 144,000 stars, GPL-3.0) is the node-based alternative, now under the Comfy-Org account after moving from comfyanonymous/ComfyUI, which still redirects. Users wire steps together as a graph to generate images, video, 3D models and audio. Its README says it can run large open models on as little as 4GB of video memory and 8GB of system memory, and it supports NVIDIA, AMD, Intel and Apple silicon. A desktop app is available for Windows and macOS.

Gemini CLI (about 109,000 stars, Apache 2.0) changed audience this year. Its README says that since 18 June 2026 free individual users and Google AI Pro and Ultra subscribers have moved to Antigravity CLI, leaving Gemini CLI for Gemini Code Assist Standard and Enterprise customers. Antigravity is the Google coding agent TSN covered when it arrived in AI Studio in March.

Claude Code (about 154,000 stars) is Anthropic’s coding agent for the terminal, IDEs and GitHub. It reads a codebase, runs routine tasks, explains code and handles git work from plain-language instructions. Its README now marks installation through npm as deprecated in favour of a native installer, Homebrew or WinGet. Its repository is public, but its licence reads “All rights reserved” and points to Anthropic’s commercial terms, so it is not open source.

RAGFlow (about 93,000 stars, Apache 2.0) is a retrieval-augmented generation engine built around deep document understanding, designed to pull knowledge out of unstructured files with complicated formats. In August it added “agentic” retrieval with adjustable thinking modes, and it released its first 1.0 candidate on 29 September 2026.

Five of the 20 come with licence strings attached

For a hobby project, the licence column hardly matters. For a company putting these tools into a product, it decides what is allowed. Dify needs a commercial licence for multi-tenant hosting, Open WebUI restricts removing its branding at scale, n8n’s licence limits commercial redistribution, AutoGPT’s platform code sits under Polyform Shield, and Stable Diffusion WebUI’s AGPL requires source sharing for networked use. Claude Code is in a different category again: a public repository is not the same as an open-source one. None of this makes the tools worse. It means “open source” on a slide is not a licence review, and the terms have shifted for several projects since they first became popular.

Stars measure attention, not upkeep

Star counts only go up. Stable Diffusion WebUI and DeepSeek-V3 both sit above 100,000, yet neither is where current development happens. Stars also say nothing about safety. JFrog’s finding that an npm worm campaign armed 42 GitHub repositories from 11 owners is a reminder that cloning and running a repository, especially inside an AI coding assistant, is code execution. The useful checks are the last commit date, the release history, the owner account and the licence file. For learning resources rather than tools, TSN’s list of AI learning repositories on GitHub covers courses and notebooks.

Next year’s version of this list will probably look different again, and the licence files are as likely to be the reason as the star counts.

20 AI GitHub Repositories, Rechecked for Licences and Upkeep

Where to start depends on the job

Agents and automation

Running models on your own hardware

Frameworks and models

Images, coding and retrieval

Five of the 20 come with licence strings attached

Stars measure attention, not upkeep

Sources

Related stories

Bulgarian Teen Held Over BNT and BNR Hacks as Defence Disputes Link

Canada Becomes EuroHPC’s 39th Member, Gaining a Path to JUPITER

Nesformer Compiles an NES Processor Into a Qwen-Style AI Model

Wiz Counts Nine AI Developer Platform Breaches in ‘Black October’

Latest stories