GitHub Trending Weekly Digest — March 9-14, 2026

Another week, another wave of innovation on GitHub. This week's trending repos span from AI agent orchestration to 1-bit LLM inference, from multi-agent simulation engines to hyper-lightweight browsers. Let's dive into what caught developers' attention from March 9-14, 2026.

🔥 Top Trending: Held the Spotlight for 3 Days

MiroFish — Multi-Agent Prediction Engine

🔗 github.com/666ghj/MiroFish
Trending: Mar 9, 10, 11

What it does: A next-generation AI prediction engine built on multi-agent technology. Upload seed materials (breaking news, policy drafts, financial signals) and MiroFish constructs a high-fidelity parallel digital world populated by thousands of independent agents with unique personalities, memories, and behavioral logic. Watch them interact, evolve socially, and predict future outcomes.

Why it matters: Traditional prediction methods rely on statistical models or single-agent inference, failing to capture emergent group dynamics and complex social interactions. MiroFish enables decision-makers to rehearse policies in a zero-risk environment and lets individuals explore creative scenarios — from predicting social sentiment to testing novel endings for stories.

Tech: Python, Node.js, GraphRAG, Zep Cloud (memory management), OpenAI-compatible LLM APIs (recommends Qwen-Plus), OASIS simulation engine (CAMEL-AI), dual-platform parallel simulation, Docker deployment

OpenRAG — All-in-One RAG Platform

🔗 github.com/langflow-ai/openrag
Trending: Mar 12, 13, 14

What it does: A comprehensive retrieval-augmented generation (RAG) platform built on Langflow, Docling, and OpenSearch. Combines document ingestion, intelligent parsing, semantic search, and AI chat in a single package with a drag-and-drop workflow builder.

Why it matters: Building enterprise RAG systems is complex, requiring integration of multiple components. OpenRAG delivers a production-ready solution out of the box: pre-packaged core tools (no separate configuration needed), agentic RAG workflows with re-ranking and multi-agent coordination, enterprise-grade search powered by OpenSearch, and both Python and TypeScript SDKs. It even includes MCP server support for Cursor/Claude Desktop integration.

Tech: FastAPI, Next.js, OpenSearch, Langflow (workflow orchestration), Docling (document parsing), Python SDK, TypeScript/JavaScript SDK, Model Context Protocol (MCP), Docker/Podman deployment

🌟 Strong Performers: Trending for 2 Days

Google Cloud Generative AI — Official Gemini Samples

🔗 github.com/GoogleCloudPlatform/generative-ai
Trending: Mar 9, 10

Google Cloud's official repository showcasing how to use Gemini and other generative AI models on Vertex AI. Features comprehensive Jupyter notebooks and code samples covering RAG, Grounding, Imagen (image generation/editing), Chirp (speech recognition), and the Agent Development Kit (ADK). An essential resource for developers entering the Google Cloud AI ecosystem.

Tech: Jupyter Notebook, Python, Vertex AI SDK, Gemini API, Imagen (vision), Chirp (audio), Agent Development Kit

Agency Agents — Production-Ready AI Agent Team

🔗 github.com/msitarzewski/agency-agents
Trending: Mar 10, 11

A complete collection of 112+ professional AI agents spanning engineering, design, marketing, product, and QA. Each agent has a unique personality, workflow, and verifiable deliverables. No manual activation needed — agents auto-trigger relevant skills. Unlike generic prompt templates, these are battle-tested specialist systems with clear success metrics. Works across Claude Code, Cursor, Aider, Windsurf, and more.

Tech: Shell scripts, Markdown agent definitions, multi-tool converter system (Claude Code native, Antigravity, Gemini CLI, OpenCode, Cursor, Aider, Windsurf)

Promptfoo — LLM Testing & Red Team Tool

🔗 github.com/promptfoo/promptfoo
Trending: Mar 10, 11

Professional testing and security toolkit for LLM applications. Supports automated evaluation of prompts, agents, and RAG systems, plus AI red team penetration testing and vulnerability scanning. Compare models (GPT, Claude, Gemini, Llama, etc.), run automated assessments, simulate red team attacks, and scan for vulnerabilities. Runs 100% locally — prompts never leave your machine. CI/CD integration and code scanning included.

Tech: TypeScript, CLI tool, declarative config files, multi-provider support, CI/CD integration, security vulnerability scanning framework

Superpowers — Agent-Driven Development Framework

🔗 github.com/obra/superpowers
Trending: Mar 11, 13

A complete software development workflow framework driven by composable "skills." Enforces TDD, YAGNI, DRY, and other best practices throughout the process: brainstorming → spec review → implementation plan → subagent-driven development (SDD). Each task is executed by independent subagents and undergoes two-stage review (spec compliance + code quality). Prevents agents from jumping straight into coding without proper design and testing.

Tech: Shell, Markdown skill definitions, Git worktrees workflow, supports Claude Code, Cursor, Codex, OpenCode, Gemini CLI, subagent orchestration system

Fish Speech — SOTA Open-Source TTS

🔗 github.com/fishaudio/fish-speech
Trending: Mar 11, 12

The most advanced open-source text-to-speech system. Fish Audio S2 model trained on 10 million+ hours of audio across ~50 languages, combining reinforcement learning alignment with dual-autoregressive architecture to generate natural, realistic, emotionally rich speech. Supports inline prosody and emotion control via natural language tags ([laugh], [whispers], [super happy]). Achieves lowest WER (word error rate) on Seed-TTS Eval benchmark (0.54% Chinese, 0.99% English), outperforming all closed-source systems.

Tech: Python, Transformer decoder-only architecture, RVQ audio codec (10 codebooks), Dual-AR architecture, GRPO (Group Relative Policy Optimization) reinforcement learning, SGLang inference optimization, 50+ language support

BitNet — 1-bit LLM Inference Framework

🔗 github.com/microsoft/BitNet
Trending: Mar 12, 13

Microsoft's official inference framework for 1-bit large language models. Achieves 1.37-5.07x speedup on ARM CPUs with 55%-70% energy reduction; 2.37-6.17x speedup on x86 CPUs with 71%-82% energy reduction. Run 100B parameter models on a single CPU at human reading speed (5-7 tokens/s). Dramatically improves local device LLM viability.

Tech: Python, C++, LLVM/Clang, llama.cpp framework, T-MAC lookup table, V8 snapshot, cmake, CUDA/GPU kernels, GGUF model format

💡 One-Day Stars: Single-Day Highlights

OpenClaw — Cross-Platform Personal AI Assistant

🔗 github.com/openclaw/openclaw
Trending: Mar 9

Local-first, always-on personal assistant that runs on any device. Supports 20+ messaging platforms (WhatsApp, Telegram, Slack, Discord, etc.), integrates voice wake-up, Canvas visualization interface, and browser control. No cloud dependency required for ChatGPT-level interaction.

Tech: Node.js (≥22), TypeScript, WebSocket, Playwright, multi-agent routing, TTS/STT, Docker sandbox, Tailscale

nanochat — Train GPT-2 for $100

🔗 github.com/karpathy/nanochat
Trending: Mar 9

Andrej Karpathy's minimalist framework for training GPT-2-level LLMs with a $100 budget. Compresses 2019's $43K GPT-2 training cost down to just $48 (≈2 hours on 8×H100). Complete pipeline: tokenization, pretraining, fine-tuning, evaluation, inference, and ChatGPT-style web UI. Automated hyperparameter calculation with a single complexity knob (--depth) lets you train and chat with your own LLM on a single GPU node.

Tech: PyTorch, BPE Tokenizer, GPT Transformer, AdamW + Muon optimizers, DCLM CORE eval, FP8/BFloat16 mixed precision, WandB

BettaFish — Multi-Platform Sentiment Analysis Assistant

🔗 github.com/666ghj/BettaFish
Trending: Mar 9

A multi-agent sentiment analysis assistant for everyone. Automatically analyzes 30+ mainstream social media platforms and millions of comments domestically and abroad. Breaks through information bubbles, restores sentiment reality, and predicts future trends. Features multi-modal capabilities (video/image parsing), agent forum collaboration, and public-private data fusion for brand reputation analysis and financial market forecasting.

Tech: Python, Flask, multi-agent (Query/Media/Insight/Report), Playwright crawler, PostgreSQL/MySQL, Qwen fine-tuned model, Streamlit, Docker

Hermes Agent — Self-Improving AI Agent

🔗 github.com/NousResearch/hermes-agent
Trending: Mar 10

The only agent system with a built-in learning loop. Creates skills from experience, self-optimizes during use, and deepens user understanding across sessions via cross-session memory. Runs on a $5 VPS or serverless architecture with near-zero cost. Supports Telegram, Discord, Slack, WhatsApp, Signal, and CLI.

Tech: Python, Node.js, FTS5 session search, Honcho dialectic user modeling, AgentSkills.io standard, Atropos RL environments, multi-provider support

InsForge — AI Coding Agent Backend Platform

🔗 github.com/InsForge/InsForge
Trending: Mar 12

A backend development platform built specifically for AI coding agents. Provides a semantic layer that exposes database, auth, storage, and functions to agents, enabling them to understand, reason, and operate backend systems end-to-end. Agents can retrieve backend context docs, directly configure backend primitives, and inspect state/logs via structured schemas. One-click Docker deployment, 2-line MCP server integration.

Tech: TypeScript, Docker, PostgreSQL, S3-compatible storage, Edge Functions, Model Gateway (OpenAI-compatible API across LLM providers), MCP (Model Context Protocol), Railway/Zeabur/Sealos one-click deploy

Hindsight — Agent Memory That Learns

🔗 github.com/vectorize-io/hindsight
Trending: Mar 12

An agent memory system that focuses on learning, not just remembering. Uses biomimetic data structures to organize memories (world facts, experiences, mental models), mimicking how human memory works. Achieves state-of-the-art performance on LongMemEval benchmark, outperforming RAG and knowledge graphs. 2-line LLM wrapper integration with automatic memory storage and retrieval. Offers three core operations: Retain (store info), Recall (retrieve via semantic/keyword/graph/temporal strategies), Reflect (deep analysis of existing memories).

Tech: Python, TypeScript, Docker, PostgreSQL, vector search (sparse/dense), BM25 keyword matching, knowledge graph (entity/temporal/causal), cross-encoder re-ranking, RRF (Reciprocal Rank Fusion), supports OpenAI/Anthropic/Gemini/Groq/Ollama/LMStudio

Lightpanda Browser — Ultra-Lightweight Headless Browser

🔗 github.com/lightpanda-io/browser
Trending: Mar 13

A headless browser built from scratch (not based on Chromium/WebKit) specifically for automation. Uses 1/9 the memory of Chrome, executes 11x faster, starts in seconds. Supports Puppeteer, Playwright, chromedp, and other mainstream protocols. Written in Zig with V8 for JavaScript execution.

Tech: Zig, V8, libcurl, html5ever, CDP (Chrome DevTools Protocol), Docker

Public APIs — Massive Free API Directory

🔗 github.com/public-apis/public-apis
Trending: Mar 13

Community-maintained directory of thousands of free public APIs spanning 90+ categories. Provides structured index including auth methods, HTTPS support, and CORS status. Covers animals, anime, blockchain, business, cryptocurrency, weather, machine learning, and more. 250K+ GitHub stars, continuously updated by the community.

Tech: Markdown, JSON, REST APIs, community-curated database

OpenViking — Context Database for AI Agents

🔗 github.com/volcengine/OpenViking
Trending: Mar 14

An open-source context database designed specifically for AI agents. Unifies management of memories, resources, and skills via a file system paradigm. Implements layered context delivery (L0/L1/L2 on-demand loading), directory recursive retrieval, visualization of retrieval trajectories, and automatic session management. Lets developers build agent "brains" like managing local files.

Tech: Python, Go, C++ compiler, Viking URI protocol, LCM transports, PaCMAP projection, Volcengine Doubao/OpenAI/LiteLLM multi-model support, Anthropic/DeepSeek/Gemini/vLLM/Ollama

Claude Plugins Official — Curated Plugin Directory

🔗 github.com/anthropics/claude-plugins-official
Trending: Mar 14

Anthropic's official high-quality Claude Code plugin directory. Provides both internal plugins (developed by Anthropic team) and vetted third-party plugins. Users can install directly via /plugin install {plugin-name}@claude-plugin-directory. Every plugin follows standard structure (plugin.json, MCP server config, commands/agents/skills) ensuring quality and compatibility.

Tech: Claude Code Plugin System, MCP (Model Context Protocol), JSON configuration, slash commands, agent definitions, skill definitions

Dimensional OS — Modern Robot OS

🔗 github.com/dimensionalOS/dimos
Trending: Mar 14

A modern operating system for general-purpose robots, no ROS required. Build physical applications in pure Python with native agent support for natural language programming and multi-agent systems. Simple installation (no ROS), next-gen SDK supporting humanoid/quadruped/drone robots. Built-in SLAM, dynamic obstacle avoidance, path planning, autonomous exploration, 3D perception, VLM, spatiotemporal RAG, object localization. Agents run as native modules, subscribe to any embedded stream (lidar, camera, control loops, motor drivers), control robots via natural language "vibecode."

Tech: Python 3.12, NixOS flakes, CUDA support, Docker, OpenSearch, Langflow, Docling, MuJoCo simulation, LCM transports, WebRTC, Ollama local LLM, Rerun visualization

Heretic — Automated Model Uncensoring Tool

🔗 github.com/p-e-w/heretic
Trending: Mar 14

Fully automated tool for removing safety alignment ("censorship") from transformer language models. Combines advanced directional ablation with TPE parameter optimization. No expensive post-training required. Users don't need to understand transformer internals — just run the CLI and the model is "uncensored." Supports most dense models, multimodal models, and some MoE architectures. Benchmarks show Heretic-generated models outperform manually tuned versions in both refusal suppression and capability retention.

Tech: Python 3.10+, PyTorch 2.2+, Optuna (TPE optimizer), bitsandbytes quantization, PaCMAP projection, Hugging Face Transformers, directional ablation/abliteration

🎯 This Week's Themes

AI Infrastructure Evolution

1-bit quantization making LLMs runnable on consumer CPUs (BitNet)
Context databases solving agent memory fragmentation (OpenViking)
Learning agents with self-improvement loops (Hermes, Hindsight)

Production-Ready Agent Frameworks

Specialized teams replacing generic prompts (Agency Agents)
Workflow orchestration enforcing best practices (Superpowers)
Backend primitives exposed for agent reasoning (InsForge)

Multi-Agent Simulation & Prediction

Emergent dynamics from thousands of interacting agents (MiroFish)
Real-world sentiment analysis across 30+ platforms (BettaFish)

Developer Productivity Wins

RAG platforms shipped as single packages (OpenRAG)
Testing frameworks for prompt quality and security (Promptfoo)
Lightweight browsers for automation (Lightpanda)

💬 Takeaway

This week's trending repos signal a maturation in the AI developer ecosystem. We're moving beyond "can we build it?" to "how do we ship it reliably, efficiently, and safely?" The emphasis on testing (Promptfoo), learning (Hermes/Hindsight), structured workflows (Superpowers), and production packaging (OpenRAG) reflects a community ready to take AI agents from demos to deployment.

The most persistent trends — MiroFish and OpenRAG — both solve "last-mile" problems: MiroFish turns speculation into simulation, OpenRAG turns documents into actionable knowledge. Meanwhile, BitNet's 1-bit inference and Lightpanda's ultra-lightweight browser show that efficiency still matters when scaling AI.

Whether you're orchestrating agent teams, building RAG pipelines, or running LLMs on consumer hardware, this week's repos offer practical tools to turn AI ambitions into shipping code.

Compiled by Tommy Zhang | March 15, 2026