GitHub trending repos for April 10, 2026 show a clear shift in developer priorities: persistent AI agents, on-device language models, and structured coding workflows dominated the weekly chart, with one non-AI tool breaking through the noise.

On-Device AI Takes Center Stage

Google’s artificial intelligence reference app, google-ai-edge/gallery, topped the list with 20,100 total stars and 4,326 new stars this week. The Kotlin-based app runs Gemma 4 and other open-source large language models directly on Android 12+ and iOS 17+ devices, with no internet connection required. Built-in features include AI chat with a “thinking mode,” camera image analysis, audio transcription, and mobile actions powered by FunctionGemma, a 270-million-parameter model.

GitHub Trending Repos Highlight Agent-Native Education

HKUDS/DeepTutor, from Hong Kong University, reached 15,700 stars after gaining 3,233 in a single week. The Python-based platform offers five learning modes: Chat, Deep Solve, Quizzes, Deep Research, and Math Animator, all within one persistent thread. It crossed 10,000 stars in 39 days and supports more than 25 large language model providers. The Math Animator feature converts abstract concepts into animated storyboards, while the Co-Writer tool embeds AI directly into a Markdown editor.

Hermes Agent Leads the Week With Nearly 20,000 New Stars

The single largest jump on the chart belonged to NousResearch/hermes-agent, which added 19,765 stars in one week, bringing its total to 50,400. The Python-based assistant runs across CLI, Telegram, Discord, Slack, and WhatsApp, and builds a persistent user profile across sessions rather than resetting after each conversation. It supports more than 200 models via OpenRouter and includes over 40 built-in tools. The project has approximately 1,000 open issues and 1,700 pull requests, indicating active community development.

Meanwhile, multica-ai/multica added 3,201 stars to reach 5,400 total. The TypeScript platform integrates coding agents such as Claude Code, Codex, and OpenCode into standard project management workflows, including kanban boards and issue trackers. Its architecture uses a Next.js frontend, a Go backend with WebSockets, and PostgreSQL storage.

Non-AI Tool Breaks Into the Top Five

siddharthvaddem/openscreen gained 12,278 stars in a week, reaching 27,400 total, making it the only repository on the list unrelated to AI. The free, open-source screen recording tool targets users of Screen Studio, a paid product priced at $29 per month. OpenScreen supports window and full-screen recording, zoom, microphone and system audio capture, cropping, trimming, annotations, and multiple aspect ratios. It runs on macOS 13+, Windows, and Linux, and carries MIT-compatible licensing that permits commercial use.

Two developer productivity repositories also drew significant attention. forrestchang/andrej-karpathy-skills, a single CLAUDE.md file encoding Andrej Karpathy’s four principles for improving LLM code output, reached 11,400 stars with 2,230 added this week. The principles cover avoiding silent assumptions, keeping code minimal, making only necessary changes, and defining verifiable success criteria before writing any code. Separately, Yeachan-Heo/oh-my-codex gained 9,737 stars in one week, reaching 20,400 total. The TypeScript project adds structured workflow commands, persistent session logs, and multi-agent coordination on top of OpenAI’s Codex CLI.

NVIDIA’s personaplex reached 8,900 stars with 2,745 added this week. The Python-based, full-duplex speech model supports 16 pre-packaged voices and allows text role prompts to switch the model’s persona between a teacher, support agent, or casual conversational partner. Model weights are available on Hugging Face under NVIDIA’s Open Model license, while the code carries an MIT license. The project is described as a research release, with production deployments requiring additional engineering work.

Rounding out the list, onyx-dot-app/onyx added 5,556 stars to reach 26,300 total. The self-hostable AI platform offers more than 50 data connectors, agentic retrieval-augmented generation, sandboxed Python code execution, voice and vision capabilities, and enterprise features including SSO, role-based access control, and audit logs. The project claimed the top position on an open Deep Research leaderboard as of February 2026.

Four Patterns Emerging Across the Weekly Chart

Across the nine repositories, four themes repeat. On-device inference is gaining traction, with Google’s Gallery app serving as a reference point for what offline AI can do on consumer hardware. Agent-native design, where systems build persistent memory and skills rather than responding to one-off prompts, appears in DeepTutor, Multica, and Hermes. Structured coding workflows, rather than prompt engineering alone, drive interest in Oh My Codex and the Karpathy-skills file. Finally, self-hosted infrastructure is a consistent preference, with Onyx, Hermes, Multica, and Google Gallery all offering options to run entirely on private hardware.