GitHub Open Source Weekly 2026-08-05: AI Learning, Agent Skills, and Human-Agent Workspaces

Data period: 2026-07-28 to 2026-08-05 (rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia

TL;DR: The fastest-growing project this week is not a new agent framework, but Microsoft's AI beginner curriculum, which added 7,554 stars. Buzz brings humans and agents into the same workspace, while projects such as book-to-skill and Open Code Review turn knowledge and engineering rules into reusable workflows. The new-repository chart is led by qm, a multiplayer agent harness, suggesting that collaboration and handoffs are becoming as important as raw model capability.

📈 Fastest Growing: Top 14

#	Project	+Stars/week	Total stars	Language	Created
1	microsoft/AI-For-Beginners	+7,554	61,640	Jupyter Notebook	2021-03-03
2	block/buzz	+7,372	22,498	Rust	2026-03-06
3	zhaoxuya520/reverse-skill	+6,154	17,789	PowerShell	2026-05-13
4	virgiliojr94/book-to-skill	+5,405	16,379	Python	2026-05-01
5	ayghri/i-have-adhd	+5,012	16,765	Python	2026-05-13
6	alibaba/open-code-review	+3,881	18,771	Go	2026-05-18
7 🔁	1jehuang/jcode	+3,735	15,869	Rust	2026-01-05
8	different-ai/openwork	+3,429	20,911	TypeScript	2026-01-14
9	moeru-ai/airi	+2,978	46,860	TypeScript	2024-12-01
10	pascalorg/editor	+2,953	21,044	TypeScript	2025-10-16
11	opengeos/GeoLibre	+2,630	5,376	TypeScript	2026-05-27
12	citrolabs/ego-lite	+2,625	8,325	JavaScript	2026-04-16
13	TencentCloud/TencentDB-Agent-Memory	+2,463	13,490	TypeScript	2026-04-07
14	lyogavin/airllm	+2,410	28,318	Jupyter Notebook	2023-06-12

🆕 Top New Repos: Top 15

#	Project	Total stars	Language	Created
1	yc-software/qm	11,109	TypeScript	2026-07-29
2	trycompai/crm	4,606	TypeScript	2026-07-31
3	bashalarmistalt/decimen-optical-transfer	4,532	TypeScript	2026-07-30
4	xdash/FDE-the-Guidance-Book-of-Forward-Deployed-Engineer	2,876	—	2026-07-30
5	FareedKhan-dev/kimi-k3-in-c	1,978	C	2026-08-01
6	microsoft/skill-recorder	1,757	TypeScript	2026-07-29
7	sqliteai/waste	1,669	C	2026-07-28
8	imsai-sh/zhuzhiliao	1,635	HTML	2026-08-01
9	DannyMac180/sol-advisor	1,299	Shell	2026-08-01
10	genspark-ai/genoffice	1,296	TypeScript	2026-07-31
11	firecrawl/anydoc	1,215	Rust	2026-08-03
12	thebuggeddev/anatomy	1,191	TypeScript	2026-08-02
13	WilonityDev/WilonityLoader	900	—	2026-07-31
14	xikhar/persona	832	JavaScript	2026-07-28
15	x4gKing/3x-ui-multi	693	JavaScript	2026-08-01

This week's highlights: Fastest Growing

1. Microsoft AI for Beginners: a structured way to learn the basics

+7,554 stars this week | 61,640 total | MIT

Microsoft's 12-week, 24-lesson curriculum covers neural networks, computer vision, NLP, generative models, and AI ethics, with notebooks, labs, and translations. For someone learning AI outside a formal course, a sequence like this is more actionable than a pile of disconnected prompts.

2. Block Buzz: a shared workspace for humans and agents

Buzz describes itself as a self-hostable workspace where humans and AI agents share rooms. Its official overview includes channels, workflows, code patches, reviews, canvases, and audit trails, with agents using their own keys and scoped memberships. A related Show HN post reached 21 points and 7 comments.

3. reverse-skill and book-to-skill: reusable knowledge as a skill

reverse-skill packages authorized reverse-engineering and security-research workflows. book-to-skill turns technical PDFs or EPUBs into an on-demand skill with chapters, a glossary, patterns, and a cheatsheet. Both point to the same shift: a skill can encode a repeatable process or a body of knowledge, not just a prompt.

4. i-have-adhd: make the answer easier to find

This project focuses on a very human problem: an agent may answer correctly while burying the answer in a wall of text. Its ADHD-friendly output approach is a reminder that usability depends on the shape of the response, not only on model quality.

5. Alibaba Open Code Review: constrain the parts that must not drift

Open Code Review combines deterministic pipelines with an LLM agent. File selection, grouping, rule matching, and comment positioning stay in the engineering layer, while the agent handles dynamic reasoning. That is a practical pattern for reliable automation: enforce invariants in code instead of hoping a prompt will preserve them.

6. jcode, openwork, and airi: different kinds of agent environments

jcode focuses on a RAM-efficient Rust harness and is the only Fastest Growing project also present in the monthly chart. openwork presents an open-source alternative to Claude Cowork, while airi explores a self-hosted AI companion with real-time voice interaction and game worlds. They serve different needs, but all move beyond a single chat box.

7. GeoLibre, ego-lite, TencentDB-Agent-Memory, and AirLLM

GeoLibre brings an open-source GIS platform into the chart. ego-lite is worth watching, although the fetched official summary does not provide enough detail for a stronger claim. TencentDB-Agent-Memory points toward database-backed agent memory. AirLLM focuses on running 70B inference on a single 4GB GPU; a related HN discussion reached 228 points and 83 comments, so the practical constraints are worth reading before trying it.

This week's highlights: Top New Repos

qm leads the new-repository chart

yc-software/qm is described as a multiplayer agent harness for work. Its HN discussion reached 672 points and 165 comments, making it the clearest signal this week that people care about coordination, handoffs, and shared state between agents.

CRM and Decimen Optical Transfer

trycompai/crm presents an open-source, agentic-first CRM and had a related HN discussion with 13 points and 8 comments. Decimen Optical Transfer uses fountain-coded QR file transfer, an interesting option when devices are isolated or a network is inconvenient. Its HN discussion was small, at 4 points and no comments, so treat it as an experiment rather than a mature transport layer.

The rest of the new chart

The remaining projects span an FDE guidance book, a C implementation of Kimi K3, Microsoft's skill-recorder, SQLite experiments, web projects, Solana tooling, GenOffice, AnyDoc, anatomy, a loader, a voice project, and a multi-instance 3x-ui project. Their data maturity varies, so the sensible next step is to read each README and recent commit before investing time.

Monthly trend comparison

Only jcode overlaps between this week's Fastest Growing list and the monthly chart. That is a conservative but useful signal: a one-week spike may follow a single exposure, while repeated presence deserves closer observation. Other monthly names still point toward agent infrastructure, design tools, office automation, and local workflows.

Three broader signals

1. Agents are becoming workspaces, not just assistants

Buzz, qm, openwork, and jcode all address some combination of task splitting, handoff, state, or traceability. The next practical improvement may be an agent that behaves like a teammate with scoped access and history, rather than another chat window.

2. Skills are becoming a format for knowledge and habits

reverse-skill, book-to-skill, i-have-adhd, and skill-recorder show skills expanding into security processes, books, output preferences, and personal workflows. That is useful, but it also makes source freshness, permissions, and maintenance important.

3. Adoption still depends on lower barriers and verifiability

AI-For-Beginners lowers the learning barrier, AirLLM targets the hardware barrier, and Open Code Review targets the verification barrier. The common thread is practical adoption: learn it, run it, and check what it did.

Conclusion

Start with AI-For-Beginners if you want a structured introduction. Look at book-to-skill if you repeatedly revisit technical books. If you are building agent workflows, Buzz, qm, and Open Code Review are more useful references than a model leaderboard alone. AirLLM and Decimen Optical Transfer are good experiments for constrained hardware or connectivity, as long as you understand their limits first.