The open-source AI projects going viral on GitHub in early November 2026 have one thing in common: they make AI agents do more on hardware you already own. One lets agents operate any desktop app, two run trillion-parameter models on an ordinary laptop, and several turn a single coding assistant into a whole team. Here are the eight worth knowing, with star counts checked directly on GitHub on November 8.
1. agent-desktop: AI agents that can use any app
vercel-labs/agent-desktop • Rust • Apache-2.0 • about 19,700 stars
The fastest climber of the week, adding thousands of stars a day. It gives AI agents "computer use" on macOS, Windows and Linux without screenshots. Instead of looking at pixels, it reads each app’s accessibility tree, the hidden map of buttons and fields that screen readers use, and hands it to the agent as a list it can click or type into. That is faster and more reliable than screenshot-based approaches. Built by Vercel Labs. Its own README warns that snapshots can capture private desktop content, so keep them out of public logs.
2. colibri: huge models on a normal computer
JustVugg/colibri • C • Apache-2.0 • about 49,100 stars
Big "mixture-of-experts" models only use a small slice of themselves for each word they write. Colibri keeps that active slice in memory and streams the rest from your SSD as needed. The result: models with hundreds of billions of parameters, including GLM-5.3, Kimi K3, DeepSeek V4.1 Flash and Mistral Large 4, running on a machine with as little as 8 GB of RAM, faster on a GPU. Pure C, no dependencies.
3. kimi-k3-in-c: a 2.78-trillion-parameter model on a CPU
FareedKhan-dev/kimi-k3-in-c • C99 • Apache-2.0 • about 11,300 stars
The headline stunt: Moonshot’s Kimi K3, one of the largest open models, running on a single CPU in 8.24 GB of RAM, with no GPU and no libraries. Read the fine print, though. The model still needs a 1.56 TB file on disk, and on an 8 GB laptop it reportedly takes around 26 seconds per token. It is a brilliant proof of what is possible, not a practical chatbot.
4. gstack: Claude Code as a full company
garrytan/gstack • TypeScript • MIT • about 140,800 stars
Y Combinator CEO Garry Tan’s personal Claude Code setup: 23 slash-command roles such as CEO, designer, engineering manager, reviewer, QA and release manager. Each one is a structured prompt and toolset, so you can ask your coding agent to "review this as a security engineer" and get a consistent process. It installs into other coding agents too, but Claude Code is the only fully supported one.
5. claude-mem: memory that carries between sessions
thedotmack/claude-mem • TypeScript • Apache-2.0 • about 114,900 stars
Coding agents forget everything when a session ends. claude-mem records what your agent did, compresses it with AI, and feeds the relevant bits back next time, so you stop re-explaining your project every morning. Despite the name, it now works with other agents too.
6. OpenMontage: a video studio run by agents
calesthio/OpenMontage • Python • AGPL-3.0 • about 78,700 stars
Turns your coding assistant into a video production team. You describe the video in plain language and the agent handles research, script, assets, editing and the final render, using generators like Veo and Kling or free stock footage. Note the AGPL licence, which has strict sharing rules if you build a commercial service on it.
7. spec-kit: GitHub’s official way to brief AI agents
github/spec-kit • Python • MIT • about 148,700 stars
GitHub’s own toolkit for "spec-driven development": write a clear specification first, then let the agent build from it. It works with Copilot, Claude Code, Cursor, Codex and others, and is trending again as teams try to make agent output more predictable.
8. auto-ontology: NVIDIA’s fix for AI guessing at your data
NVIDIA/auto-ontology • Python • Apache-2.0 • about 10,400 stars
Asking an AI questions about company databases often produces confident wrong answers. NVIDIA’s framework builds a map of what your data means, then generates queries from that map so answers are repeatable and traceable to their source. It is aimed at enterprises and uses NVIDIA’s Nemotron models by default.
What do these trends tell us?
Three things. Agents are moving off the browser and onto the whole computer. Running giant models locally is becoming realistic, which matters for privacy and cost and adds to the shift towards open models. And developers want structure, not just raw power: roles, specs and memory that make agents behave the same way twice.
Are viral GitHub projects safe to install?
Not automatically. Stars measure popularity, not security. October showed how attackers target exactly these tools, from a fake Claude Code extension to a poisoned LiteLLM release. Install from the official repository, check who maintains it, read what permissions it asks for, and be extra careful with anything that controls your desktop or reads your files.
