Top 10 GitHub Repos of the Week
The week of 2026-05-03 to 2026-05-09 was dominated by substrate projects for AI agents — virtual filesystems, inference engines, evaluation harnesses, and memory toolkits — rather than yet another…
- Click the button to download the
-bookmarks.htmlfile. - Chrome / Edge / Brave: Bookmarks → Import bookmarks and settings → Favorites or bookmarks HTML file → pick the downloaded file.
- Firefox: Bookmarks → Manage Bookmarks → Import and Backup → Import Bookmarks from HTML.
- Safari: File → Import From → Bookmarks HTML File.
Top 10 GitHub Repos of the Week
Week of: 2026-05-03 to 2026-05-09 Focus: AI tools, automations, agents, integrations, MCP, CLIs, and developer tooling. Reporting window: Sunday–Saturday of the prior calendar week (does not include the current in-progress week).
Overview
The week of 2026-05-03 to 2026-05-09 was dominated by substrate projects for AI agents — virtual filesystems, inference engines, evaluation harnesses, and memory toolkits — rather than yet another wave of monolithic frameworks. Four brand-new repos cleared 1,000+ stars in their first week (mirage, how-to-train-your-gpt, awesome-agentic-ai-zh, tokenspeed), and three more sailed past 600 (Photo-agents, opensquilla, openpets). Token efficiency, Blackwell-class inference, and the now-standard "skills + MCP + claude-code" stack ran through almost every pick — the long tail kept pushing into specialty MCP servers (ads ops, coding-agent memory) and self-evolving / vision-grounded agent research.
1. mirage
- Repo: https://github.com/strukto-ai/mirage
- Stars: 2,284 (at time of generation)
- Category: AI Tools
- Status this week: Trending this week
A unified virtual filesystem for AI agents — a sandboxable layer that agents can read, write, and snapshot against without touching the real disk. Ships adapters for LangChain, the OpenAI Agents SDK, Claude Code, and shell agents, with a TypeScript core and Python clients. Created 2026-05-06 and cleared 2,200 stars within nine days, the largest single-week surge of any AI-tooling repo in the window.
Why it's on the list: A clean "agent root filesystem" primitive plugs a gap every multi-tool agent harness has been hand-rolling.
2. how-to-train-your-gpt
- Repo: https://github.com/raiyanyahya/how-to-train-your-gpt
- Stars: 1,568 (at time of generation)
- Category: AI Tools
- Status this week: Trending this week
A from-scratch PyTorch implementation of a modern GPT-style LLM where every line is commented and the architecture (attention, tokenisation, training loop) is walked through "explained like we are five." Created 2026-05-03 and picked up more than 1,500 stars in a week as a fresh entry in the "build-an-LLM" educational canon next to nanoGPT and llm.c.
Why it's on the list: Cleanest new educational LLM build for engineers who want to understand the stack without wading through research code.
3. awesome-agentic-ai-zh
- Repo: https://github.com/WenyuChiou/awesome-agentic-ai-zh
- Stars: 1,413 (at time of generation)
- Category: AI Agents
- Status this week: Trending this week
A structured trilingual (Traditional Chinese / Simplified Chinese / English) learning roadmap for AI agents, organised as staged practicums with required readings, tagged with Claude Code, Claude Skills, MCP, and LLM-agent fundamentals. Created 2026-05-04, crossed 1,400 stars by 2026-05-15, and already has 151 forks — strong contributor velocity for a curriculum repo.
Why it's on the list: The first "awesome-list" of the agentic-AI era to ship as a proper curriculum rather than an unranked URL dump.
4. tokenspeed
- Repo: https://github.com/lightseekorg/tokenspeed
- Stars: 1,027 (at time of generation)
- Category: AI Tools
- Status this week: Trending this week
A "speed-of-light" LLM inference engine in Python targeting Blackwell-class GPUs, with kernels tuned for gpt-oss, DeepSeek, Kimi, MiniMax, and Qwen out of the box. Created 2026-05-06 and crossed 1,000 stars in nine days as the open-source inference space keeps chasing vLLM and SGLang on raw throughput.
Why it's on the list: Strongest new contender in the "fast inference for open weights" segment this week — worth watching against vLLM/SGLang benchmarks.
5. Photo-agents
- Repo: https://github.com/jmerelnyc/Photo-agents
- Stars: 852 (at time of generation)
- Category: AI Agents
- Status this week: New this week
A computer-use agent stack built around "photographic" layered memory — the agent captures vision snapshots of the screen, writes its own skills from successful trajectories, and revises them over time. Created 2026-05-04 and shipped straight into the autonomous-agent / self-evolving-agent research thread, picking up over 850 stars by week 2.
Why it's on the list: Concrete implementation of self-written skills tied to vision memory — a useful primitive for desktop and OS-control agents.
6. opensquilla
- Repo: https://github.com/opensquilla/opensquilla
- Stars: 815 (at time of generation)
- Category: AI Agents
- Status this week: New this week
A token-efficient AI agent runtime pitching "same budget, higher intelligence density" — a programming model for skills + memory tuned around minimising tokens per useful action. Created 2026-05-06; one of several "openclaw"-family entrants this week pushing the cost-efficient agent angle and already past 800 stars.
Why it's on the list: Worth tracking as cost-per-task becomes the next axis after raw capability in the agent runtime race.
7. openpets
- Repo: https://github.com/alvinunreal/openpets
- Stars: 649 (at time of generation)
- Category: Integrations
- Status this week: New this week
An Electron + Bun desktop pet that connects to Claude Code (and other coding agents) via MCP and animates its mood based on the agent's live status — running tool, waiting on user, errored, idle. Created 2026-05-04 with a companion repo alvinunreal/claude-pets (now archived) that handles the Claude Code hook side.
Why it's on the list: Best example this week of the "ambient awareness for long-running coding agents" UX problem getting a playful native solution.
8. agent-skills-eval
- Repo: https://github.com/darkrishabh/agent-skills-eval
- Stars: 492 (at time of generation)
- Category: AI Tools
- Status this week: New this week
A CLI test runner for agent skills in the agentskills.io style — feed it YAML/JSONL skill definitions plus eval cases and it scores them against any OpenAI-compatible endpoint. Created 2026-05-06; lands as the agentic-skills ecosystem matures past "marketplace of YAMLs" into "marketplace plus eval harness."
Why it's on the list: Fills the eval gap for skill-based agents the same way inspect_ai filled it for general LLM evals.
9. Agent_Memory_Techniques
- Repo: https://github.com/NirDiamant/Agent_Memory_Techniques
- Stars: 318 (at time of generation)
- Category: AI Agents
- Status this week: New this week
A 30-notebook cookbook covering conversation buffers, vector stores, knowledge graphs, episodic vs. semantic memory, MemGPT, Mem0, Letta, Zep, and Graphiti, with LoCoMo-benchmark code and production patterns. Created 2026-05-05 by the author of the popular Prompt_Engineering and GenAI_Agents cookbooks.
Why it's on the list: Best single-source reference for the agent-memory toolchain that exploded over the last six months — directly runnable Jupyter, not just diagrams.
10. whale
- Repo: https://github.com/usewhale/whale
- Stars: 123 (at time of generation)
- Category: CLI
- Status this week: New this week
A Go-based DeepSeek-native terminal coding agent that ships MCP support, agent skills, shell tools, and prefix-cache optimisation out of the box — effectively a DeepSeek-first answer to Claude Code's CLI. Created 2026-05-06; small but well-scoped and uses the modern coding-agent feature set rather than reinventing it.
Why it's on the list: First credible terminal-native coding agent built around DeepSeek as the primary model rather than as an afterthought provider.
Honorable Mentions (Optional)
- KevRojo/Dulus — Open-Claw-inspired agent that proxies through web chats to cut LLM costs ~60% by routing through claude-code (224 ★).
- Naroh091/hermes-war-room — Vue UI layer for Hermes' native multi-agent orchestration features (218 ★).
- awizemann/harness — Plain-language LLM-driven UI testing for iOS Simulator, macOS apps, and web apps in Swift 6 (187 ★).
- Sayhi-bzb/Agent-HTML — Semantic HTML scaffolding designed for AI-agent collaboration, replacing long-running Markdown artifacts (177 ★).
- enmanuelmag/agent-harness-kit — Provider-agnostic scaffolding for structured multi-agent workflows with built-in OpenTelemetry (152 ★).
Sources
- mirage — https://github.com/strukto-ai/mirage
- how-to-train-your-gpt — https://github.com/raiyanyahya/how-to-train-your-gpt
- awesome-agentic-ai-zh — https://github.com/WenyuChiou/awesome-agentic-ai-zh
- tokenspeed — https://github.com/lightseekorg/tokenspeed
- Photo-agents — https://github.com/jmerelnyc/Photo-agents
- opensquilla — https://github.com/opensquilla/opensquilla
- openpets — https://github.com/alvinunreal/openpets
- claude-pets (companion, archived) — https://github.com/alvinunreal/claude-pets
- agent-skills-eval — https://github.com/darkrishabh/agent-skills-eval
- Agent_Memory_Techniques — https://github.com/NirDiamant/Agent_Memory_Techniques
- whale — https://github.com/usewhale/whale
- Dulus — https://github.com/KevRojo/Dulus
- hermes-war-room — https://github.com/Naroh091/hermes-war-room
- harness — https://github.com/awizemann/harness
- Agent-HTML — https://github.com/Sayhi-bzb/Agent-HTML
- agent-harness-kit — https://github.com/enmanuelmag/agent-harness-kit
- GitHub search — MCP topic, created window 2026-05-03..2026-05-09 — https://github.com/search?q=topic%3Amcp+created%3A2026-05-03..2026-05-09&type=repositories&s=stars&o=desc
- GitHub search — ai-agents topic, created window 2026-05-03..2026-05-09 — https://github.com/search?q=topic%3Aai-agents+created%3A2026-05-03..2026-05-09&type=repositories&s=stars&o=desc
- GitHub search — llm topic, created window 2026-05-03..2026-05-09 — https://github.com/search?q=topic%3Allm+created%3A2026-05-03..2026-05-09+stars%3A%3E50&type=repositories&s=stars&o=desc
More from GitHub Repos
