GitHub 2026-03-15 Weekly Trending Projects

MiroFish is a multi-agent, long-memory simulation prediction platform that maps real-world seed data into high-fidelity digital worlds, enabling natural-language-driven scenario simulations and automated report generation for public opinion, policy and creative sandboxing, while requiring attention to licensing and operational risks.

Persona-centric AI agent collection with production-ready Claude Code templates for rapid specialist assistants; license and maintenance transparency are lacking.

BitNet is Microsoft's open-source inference framework that uses 1-bit (1.58-bit) representations and optimized kernels to reduce energy and accelerate large-model inference on CPU/GPU, enabling efficient edge and on-device deployments.

A Google Cloud Vertex AI sample repository covering Gemini, RAG, vision and audio workflows—useful for learning and prototyping; verify license and maintenance before production use.

Hermes Agent is a server-deployable autonomous AI agent offering a terminal-first interface, multi-platform messaging gateway, persistent memory and skills hub, sandboxed terminals, and scheduling—suitable for long-running, auditable, and extensible automation in engineering and operations.

Promptfoo is a developer-focused CLI and library for automated LLM evaluation and red-teaming, supporting local/private execution, multi-model comparisons, and CI integration to improve model reliability and security.

Lightpanda is an ultra-light headless browser (CDP-compatible) focused on low memory and fast execution for large-scale scraping, LLM training and automation.

BettaFish is a modular multi-agent public-opinion analysis engine for enterprises and researchers that aggregates multi-source multimodal data and produces actionable reports via agent-forum collaboration; however, license, dependency and operational details must be clarified to reduce production risks.

An AI-driven hedge-fund proof-of-concept for research and education that composes multiple investing-persona agents, backtesting, CLI and web UI, and configurable LLM/data integrations — enabling strategy prototyping and model evaluation while explicitly disallowing real-money trading; note it lacks a declared license and production-grade guarantees.

Extensible super-agent runtime with skills, sandbox execution and memory.

Fish-Speech (now OpenAudio) centers on large-model TTS to deliver high-fidelity, low-latency multilingual synthesis suitable for research validation and constrained non-commercial deployments; however, users must evaluate weight licensing and repository maintenance consistency to mitigate legal and reproducibility risks.

notebooklm-py delivers full programmatic access to Google NotebookLM via async Python API and CLI for research and prototyping; it depends on undocumented APIs and has unclear licensing, so avoid production use.

OpenViking abstracts agent context as a manageable 'filesystem', offering tiered loading, directory-recursive retrieval and visualization to reduce retrieval cost and improve observability—suitable for AI agent systems requiring long-term memory and complex context management.

nanochat is a minimal, hackable end‑to‑end LLM training and chat framework that can rapidly reproduce GPT‑2‑grade models on a single machine or an 8×H100 node; it is well suited for research, teaching, and fast experimentation, but pay attention to license, contributor metadata, and hardware costs.