GitHub Trending Weekly Digest — April 6-10, 2026

This week's GitHub Trending roundup (April 6-10) reveals a powerful convergence: AI agents are maturing from novelty to necessity, edge AI is bringing models to your pocket, and developer tools are getting smarter about understanding code. From self-evolving assistants to tokenizer-free voice synthesis, here's what dominated the charts.

🔥 Persistent Trending Champions

Hermes Agent — The AI That Learns While You Sleep

🔗 github.com/NousResearch/hermes-agent
Trending: 3 days (April 6, 9, 10)

What it does:
A self-improving AI agent built by Nous Research that runs anywhere—$5 VPS, GPU clusters, or serverless platforms. Unlike typical chatbots that reset each conversation, Hermes builds long-term memory through a built-in learning loop. It creates skills from experience, searches its own conversation history, and constructs persistent user models across sessions.

Why it matters:
This breaks the biggest limitation of current AI assistants: amnesia. Hermes can run 24/7 on remote infrastructure (controllable via Telegram, Discord, Slack, Signal), spawn sub-agents for parallel work, and improve autonomously. It's compatible with 200+ models (OpenRouter, Nous Portal, OpenAI, Anthropic) with zero vendor lock-in. One-line install, MIT licensed, and already research-ready with support for trajectory generation and RL training.

Tech:
Python + Node.js, 6 terminal backends (local, Docker, SSH, Daytona, Singularity, Modal), FTS5 session search + LLM summaries + Honcho user modeling, built-in cron scheduler, agentskills.io standard compatible.

Google AI Edge Gallery — LLMs in Your Pocket, No Server Required

🔗 github.com/google-ai-edge/gallery
Trending: 3 days (April 6-8)

What it does:
Google's official showcase for running open-source LLMs (including Gemma 4) entirely on mobile devices. All inference happens locally—no network, no cloud, no data leaving your phone. Supports multimodal tasks: chat with Thinking Mode (watch the model reason), image recognition, voice transcription, and even natural language device control.

Why it matters:
Privacy is the killer feature. Your prompts, images, and sensitive data never leave the device. It's also a platform for developers to experience and benchmark the latest on-device AI capabilities (Gemma, Llama, Phi-4, Qwen). Agent Skills tiles let you enhance LLMs with Wikipedia lookups, maps, visualization cards, and community-contributed skills loaded from URLs or GitHub Discussions.

Tech:
Kotlin (Android 12+) + Swift (iOS 17+), LiteRT lightweight runtime, Tree-sitter AST parsing, Hugging Face integration. Open source, available on Google Play and App Store.

GitNexus — The Code Graph That Makes AI Smarter

🔗 github.com/abhigyanpatwari/GitNexus
Trending: 2 days (April 6-7)

What it does:
A zero-server code intelligence engine that indexes your entire codebase as a knowledge graph. It tracks every dependency, call chain, cluster, and execution flow—then exposes this relational intelligence to AI agents via 16 MCP tools. Think of it as pre-computed architecture awareness for Cursor, Claude Code, Codex, and Windsurf.

Why it matters:
AI coding tools are powerful but blind to codebase structure. When they modify a function, they often miss the 47 other places that depend on it. GitNexus solves this with pre-computed relational intelligence: one query returns complete context (no multi-hop RAG chains), small models get full architecture views, and you get blast radius analysis before making breaking changes. It also auto-generates Agent Skills for your codebase.

Tech:
TypeScript, Tree-sitter AST parsing, LadybugDB (embedded graph database + vector support), MCP protocol, supports 14 languages (TypeScript, Python, Java, Go, Rust, C#, Kotlin, PHP, Ruby, Swift, C, C++, Dart), Web UI with Sigma.js + Graphology for visual graph exploration.

💪 Multi-Day Contenders

LiteRT-LM — Google's Production-Grade Edge AI Framework

🔗 github.com/google-ai-edge/LiteRT-LM
Trending: 2 days (April 7-8)

What it does:
Google's open-source framework for deploying LLMs on edge devices (Android, iOS, Web, Desktop, IoT including Raspberry Pi). Optimized for peak performance via GPU/NPU accelerators, supports multimodal inputs (vision, audio), and enables function calling for agent workflows.

Why it matters:
This is the engine behind Google's production deployments—Chrome, Chromebook Plus, Pixel Watch. The LiteRT-LM CLI lets you run models directly from the terminal without writing code (litert-lm run --from-huggingface-repo gemma-4-E2B-it). Multi-language APIs (C++, Kotlin, Python, Swift) make it developer-friendly, and Gemma 4 support demonstrates state-of-the-art edge performance.

Tech:
C++ core, LiteRT runtime, hardware acceleration (GPU/NPU), Tree-sitter, one-line CLI install via uv, supports Gemma, Llama, Phi-4, Qwen with INT4 quantization.

PersonaPlex — NVIDIA's Full-Duplex Voice AI

🔗 github.com/NVIDIA/personaplex
Trending: 2 days (April 7-8)

What it does:
Real-time, full-duplex voice-to-voice conversation model from NVIDIA. You can define AI personality via text prompts ("you are a friendly teacher" or "customer service agent for waste management") and control voice timbre with audio samples. True bidirectional interaction—users can interrupt, pause, backchannel naturally.

Why it matters:
Most voice AI is half-duplex: you talk, it waits, it responds. PersonaPlex delivers human-like conversation flow with low latency, consistent character personality, and 8 natural voices + 10 diverse voices out of the box. Built on Moshi architecture, trained on synthetic dialogues and Fisher English Corpus real conversations. Suitable for QA assistants, customer service scenarios, and casual chat.

Tech:
Python, Moshi architecture, Opus audio codec, Web UI + offline scripts, supports CPU offload via accelerate, NVIDIA Open Model License.

Andrej Karpathy's Claude Code Principles

🔗 github.com/forrestchang/andrej-karpathy-skills
Trending: 2 days (April 8-9)

What it does:
A single CLAUDE.md file distilling Andrej Karpathy's observations of common LLM coding mistakes. Four principles: Think Before Coding (explicit reasoning), Simplicity First (minimal code, no speculative features), Surgical Modifications (only touch what's necessary), Goal-Driven Execution (define success, loop until tests pass).

Why it matters:
LLMs love to over-engineer. They assume, over-complicate, and touch code they shouldn't. This skill plugin (or project-level CLAUDE.md) addresses these flaws directly. Community feedback: fewer unnecessary changes, cleaner PRs, better first-time implementations. It converts vague instructions into declarative goals: instead of "add validation," you say "write tests for invalid inputs, then make them pass."

Tech:
Claude Code plugin (via /plugin marketplace add) or standalone CLAUDE.md file, MIT licensed.

SEO Machine — Claude Code for Content Marketers

🔗 github.com/TheCraigHewitt/seomachine
Trending: 2 days (April 8-9)

What it does:
A professional SEO content workspace built on Claude Code. Automates the full workflow: keyword research, competitor analysis, long-form article writing (2000-3000+ words), optimization, and WordPress publishing (with Yoast SEO metadata). Integrates Google Analytics 4, Search Console, and DataForSEO for performance-driven strategy.

Why it matters:
SEO content creation is manual, fragmented, and repetitive. SEO Machine consolidates it into command-driven workflows (/research, /write, /optimize, /publish-draft). Five Python analysis modules handle search intent classification, keyword density clustering, readability scoring, and SEO quality rating. It also includes 26 marketing skills (copywriting, CRO, A/B testing, email sequences) and supports landing page optimization.

Tech:
Claude Code + Python (FastAPI, LlamaIndex), WordPress REST API, GA4/GSC/DataForSEO integration, 26 marketing skills, Apache-2.0 licensed.

🌟 Single-Day Standouts

OpenScreen — Free Screen Studio Alternative

🔗 github.com/siddharthvaddem/openscreen
Trending: April 6

What it does: Open-source screen recorder with auto/manual zoom, system audio capture, annotations, custom backgrounds, and dynamic blur. 100% free for personal and commercial use.
Why it matters: Screen Studio costs $29/month. OpenScreen offers essential features without the subscription.
Tech: Electron + React + TypeScript + Vite, PixiJS (WebGL), cross-platform (macOS, Windows, Linux).

Shannon — Autonomous Penetration Testing

🔗 github.com/KeygraphHQ/shannon
Trending: April 6

What it does: AI-powered pen testing tool that analyzes source code, identifies attack vectors, and executes real exploits (OWASP coverage: injection, XSS, SSRF, auth/authz). Auto-handles 2FA/TOTP, browser navigation, parallel vulnerability analysis.
Why it matters: Traditional pen tests happen once a year. Shannon lets you test every build or release, filling the 364-day security gap.
Tech: TypeScript, Anthropic Claude Agent SDK, Temporal, Nmap/Subfinder/WhatWeb/Schemathesis, Playwright.

qmd — Local-First Search for Docs & Code

🔗 github.com/tobi/qmd
Trending: April 7

What it does: Mini CLI search engine combining BM25 (keyword), vector search, and LLM reranking. Fully local, designed for docs, knowledge bases, meeting notes. AST-aware chunking for code files.
Why it matters: Privacy-first search for AI agents. MCP integration with Claude Desktop, Claude Code, Cursor, Codex. Position-aware hybrid fusion with clickable terminal links.
Tech: TypeScript, Node.js ≥22, embeddinggemma-300M, qwen3-reranker, SQLite (FTS5 + sqlite-vec), multi-language support (CJK).

DeepTutor — Personalized Learning Assistant

🔗 github.com/HKUDS/DeepTutor
Trending: April 9

What it does: Agent-native learning platform with 5 modes (chat, deep solve, quiz generation, research, math animation) sharing one thread. Personal TutorBots with independent memory, personality, skills. Persistent user profile across all features.
Why it matters: Learning tools are fragmented. DeepTutor unifies them with multi-agent deep solving (plan→research→solve→verify), guided learning plans, AI Co-Writer for text refinement, and RAG-powered Knowledge Center.
Tech: Python 3.11+ (FastAPI, LlamaIndex), Next.js 16 + React 19, 30+ LLM providers (OpenAI, Anthropic, Gemini, Qwen, DeepSeek), Docker one-click deployment, Apache-2.0 licensed.

VoxCPM — Tokenizer-Free Voice Synthesis

🔗 github.com/OpenBMB/VoxCPM
Trending: April 9

What it does: VoxCPM2 generates 48kHz studio-quality audio from text in 30 languages (including Chinese dialects). Supports voice design from natural language descriptions only, controllable cloning with style guidance, and ultimate cloning (seamlessly continue from reference audio).
Why it matters: Traditional TTS uses discrete tokens, limiting naturalness. VoxCPM's end-to-end diffusion autoregressive approach achieves SOTA on Seed-TTS-eval, CV3-eval, and InstructTTSEval. RTF ~0.13 after Nano-vLLM acceleration.
Tech: 2B params (MiniCPM-4 base), 2M hours multilingual training, AudioVAE V2 asymmetric encoding/decoding, CLI tools (voxcpm design/clone/batch), Apache-2.0 licensed.

MarkItDown — Universal Document Converter

🔗 github.com/microsoft/markitdown
Trending: April 10

What it does: Microsoft's tool to convert PDF, PowerPoint, Word, Excel, images (EXIF + OCR), audio (transcription), HTML, CSV/JSON/XML, ZIP, YouTube, EPub to Markdown for LLMs.
Why it matters: Markdown is LLM-friendly, preserves structure, and has high token efficiency. Includes MCP server for Claude Desktop integration. Plugin system supports Azure Document Intelligence.
Tech: Python 3.10+, modular dependencies (pdf, docx, pptx, xlsx, audio-transcription), CLI + API + Docker.

Rowboat — Local-First AI Colleague

🔗 github.com/rowboatlabs/rowboat
Trending: April 10

What it does: AI assistant with long-term memory stored locally as Markdown (Obsidian-compatible). Connects to Gmail, Google Calendar, Google Drive, Fireflies meeting notes. Builds knowledge graph over time.
Why it matters: Traditional AI tools retrieve fresh each time. Rowboat maintains compounding context. Privacy-first: all data local, no proprietary formats. MCP protocol extends tools (Exa search, Twitter/X, Slack, Linear/Jira, GitHub).
Tech: TypeScript, local Markdown vault, Ollama/LM Studio (local models) or BYOK (hosted), Live Notes, meeting prep, email drafting, voice memos (Deepgram + ElevenLabs).

Multica — Hosted Agent Platform

🔗 github.com/multica-ai/multica
Trending: April 10

What it does: Turns coding agents (Claude Code, Codex, OpenClaw, OpenCode) into team members. Agents appear on boards, autonomously report blockers, update statuses. Every solution becomes a reusable skill (deployment, migration, code review).
Why it matters: Eliminates copy-paste prompts and babysitting. Unified management of local daemons and cloud runtimes, real-time monitoring. Multi-workspace isolation for teams.
Tech: Next.js 16 (frontend), Go (backend: Chi, sqlc, gorilla/websocket), PostgreSQL 17 + pgvector, Docker Compose self-hosted or Multica Cloud.

Archon — Workflow Engine for AI Coding

🔗 github.com/coleam00/Archon
Trending: April 10

What it does: First open-source workflow engine for AI coding agents. Encodes development processes as YAML (plan→implement→verify→review→PR). Deterministic structure, AI fills in intelligence at each step. Independent git worktrees for parallel execution.
Why it matters: AI agents are non-deterministic—Archon makes them predictable. Fire-and-forget task execution. 17 built-in workflows (issue fixes, PR reviews, feature dev, conflict resolution). Web dashboard with drag-drop DAG workflow builder.
Tech: TypeScript, YAML workflow definitions (like Dockerfile for infra, GitHub Actions for CI/CD), CLI + Web UI + Slack/Telegram/GitHub integration.

📊 This Week's Themes

Edge AI Goes Mainstream
Google dominated with two projects (AI Edge Gallery, LiteRT-LM) showcasing production-grade on-device inference. Privacy and offline capability are no longer experimental—they're shipping in Chrome, Chromebook Plus, and Pixel Watch.

AI Agents Mature from Chatbots to Colleagues
Hermes Agent, Multica, and Archon represent a shift: agents are no longer just tools you prompt—they're persistent systems that learn, collaborate, and compound skills over time. Hermes even runs 24/7 on remote infrastructure.

Code Intelligence Gets Relational
GitNexus and qmd tackle the same core problem: AI needs structured knowledge, not just vectors. Pre-computed graphs, hybrid search (BM25 + vector + reranking), and MCP integration are the new table stakes for developer tooling.

Developer Workflow Automation
SEO Machine, DeepTutor, and Archon show specialized automation replacing manual workflows. Not just "help me write code"—these tools orchestrate entire processes (research→write→optimize→publish).

Voice AI Gets Real-Time
PersonaPlex (NVIDIA) and VoxCPM (OpenBMB) both push voice synthesis and conversation beyond half-duplex turn-taking. Full-duplex, tokenizer-free, and 48kHz studio quality are the new benchmarks.

💡 Takeaway

This week's trending repos reveal a clear trajectory: AI is moving from assistive to autonomous, from cloud to edge, and from generic to workflow-specific. The projects that stayed trending longest (Hermes Agent, AI Edge Gallery, GitNexus) all share a common trait—they don't just add AI features, they rethink entire workflows around persistent intelligence.

If you're building AI tooling in 2026, the winning formula is: local-first for privacy, graph-based for structure, and persistent for compounding value.

Compiled by Tommy Zhang | April 12, 2026