GitHub 2026-03-21 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.

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

Open SWE is an open-source asynchronous coding agent that uses LLMs to plan and implement repository-wide changes—from issue-triggered tasks to automated PRs—enabling teams to scale code maintenance and workflow automation.

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.

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

Superpowers provides a composable skills library and TDD-enforced agent workflow for Claude Code, enabling end-to-end automation from design and planning to subagent execution, with plugin integration and extensibility.

Native HUD for Claude Code that visualizes context, tools and agents for terminal debugging.

DeepAgents is a Python toolkit for building layered, extensible deep agents that combine planning, subagents, filesystem access and persistent memory to handle long, complex tasks and research workflows; however, repository metadata (license, contributors, releases) is incomplete, so evaluate compliance and stability before production use.

PageAgent is an in-page natural-language agent embeddable with a single script, enabling front-ends to control DOM and forms via custom LLMs without backend changes—ideal for shipping AI copilots in SaaS and enterprise admin UIs.

A progressive, hands-on tutorial repository for developers and researchers that demonstrates how to build Claude-like AI coding agents across five increasing-complexity examples, emphasizing clear concepts and runnable samples.

DimensionalOS is a Python-centric, agentive operating system for physical-space robotics integrating SLAM, perception and spatio-temporal memory to build multi-agent applications in simulation and on hardware; however, license and maintenance records are currently incomplete, so assess compliance and long-term support before adoption.

Impeccable supplies an LLM-driven frontend design skillset—commands and references to audit, polish, and extract UI patterns—enabling teams to apply consistent design best practices across AI tools.

Everything Claude Code delivers a production-grade toolchain for agent harness performance and governance—covering skills, hooks, memory persistence, and security scanning—suited for teams requiring cross-platform parity and measurable quality gates.

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.

Context Hub: Provides Markdown‑maintained, versioned API docs and local annotation/feedback loops for coding agents, improving code-generation accuracy and continual learning.