GitHub Trending Weekly, August 19, 2026: AI Tools Add Memory, Permissions, and Everyday Access
Data period: August 14–19, 2026 Sources: GitHub Trending weekly and monthly, GitHub Search API, GitHub REST API, and HN Algolia
TL;DR: diagram-design, public-apis, and MoneyPrinterTurbo led this week's growth ranking, but the middle of the table tells the more interesting story. On-device models, agent memory, traceable knowledge, and local hardware tools all gained attention. The new-repository list brought agents into team chat, dedicated computers, and shopping decisions. Stars show where attention is moving, but they cannot tell you whether a tool is mature, especially when a project is only days old.
📈 Fastest Growing: Top 10 by stars gained
🔁 marks projects that also appeared in the monthly trend. Figures come from the snapshot saved on August 21, 2026. They are not a guarantee of quality or safety.
| # | Project | +Stars | Total stars | Language | Created |
|---|---|---|---|---|---|
| 1 | diagram-design | +11,325 | 24,731 | HTML | 2026-04-16 |
| 2 | public-apis | +11,259 | 467,632 | Python | 2016-03-20 |
| 3 | MoneyPrinterTurbo | +9,712 | 113,502 | Python | 2024-03-11 |
| 4 | semantica | +3,674 | 9,995 | Python | 2025-06-25 |
| 5 🔁 | needle | +3,409 | 8,243 | Python | 2026-02-24 |
| 6 | unsloth | +3,300 | 74,189 | Python | 2023-11-29 |
| 7 | OpenLogi | +2,674 | 12,459 | Rust | 2026-05-24 |
| 8 | OpenViking | +2,444 | 31,401 | Python | 2026-01-05 |
| 9 🔁 | omarchy | +2,395 | 27,147 | Shell | 2025-06-01 |
| 10 | ai-memory | +1,952 | 3,838 | Rust | 2026-05-21 |
🆕 Top New Repos: Top 10 created this period
| # | Project | Total stars | Language | Created |
|---|---|---|---|---|
| 1 | dsh-routing-suite | 6,521 | PowerShell | 2026-08-14 |
| 2 | dsh-anchored-standard | 3,697 | JavaScript | 2026-08-14 |
| 3 | ip-as-logo-skill | 3,389 | Not specified | 2026-08-18 |
| 4 | cumora | 2,834 | TypeScript | 2026-08-17 |
| 5 | OpenBot | 1,930 | TypeScript | 2026-08-17 |
| 6 | dsh-market | 1,578 | TypeScript | 2026-08-14 |
| 7 | northcinder | 1,203 | JavaScript | 2026-08-17 |
| 8 | zsvirt | 1,165 | Java | 2026-08-14 |
| 9 | deepseek-harness-orange-book | 1,117 | HTML | 2026-08-14 |
| 10 | Deepseek-Harness-EAC | 1,064 | JavaScript | 2026-08-14 |
Fastest Growing: Start with fit, not stars
1. diagram-design
This diagram-design skill is built for coding agents. It produces self-contained HTML and SVG files and can redraw Mermaid or draw.io sources. Its appeal is not simply that it draws diagrams. It turns brand fonts, colors, and layouts into a repeatable process. If you make presentations, tutorials, or system diagrams every week, compare the same brief in this tool and your current setup. If you only draw occasionally, inspect the examples before changing your workflow.
2. public-apis
This free API directory has existed for years and still gained 11,259 stars during the period. It is useful for finding data sources for side projects and automation prototypes. Inclusion in the directory does not mean an API is stable, its free allowance will remain available, or its data can be used commercially. Before building anything, check each provider's official documentation for authentication, quotas, and licensing.
3. MoneyPrinterTurbo
The official README combines scripts, media, subtitles, background music, and short-video assembly in one workflow. For people producing short videos in batches, the benefit is fewer tool switches. If originality matters, media rights, narration quality, and platform policies on repetitive content come first. The tool can reduce production friction, but it cannot give the content a point of view.
4. semantica
semantica builds context graphs and knowledge graphs from enterprise data while retaining the sources behind decisions. For teams that need audits, data lineage, or regulated workflows, that is closer to the real problem than plain vector search. For personal notes without complex governance needs, the cost of building and maintaining a graph may exceed the benefit. Performance and positioning claims in the README come from the project itself, so test them on your own data before adopting it.
5. needle
Needle 2 is a small model for tool calls, device operations, and structured extraction. The project says the entire model is a single 14MB binary, uses about 28MB of RAM for a full session, and can return schema-constrained tool calls without online inference. That makes it relevant to wearables, smart homes, and narrow privacy-sensitive tasks. Its small size does not make it a replacement for a general model. The related HN discussion reached 535 points and 185 comments, with much of the debate focused on whether efficiency in constrained tasks offsets the loss of general capability.
6. unsloth
Unsloth packages local model use and training behind a desktop interface. The GUI lowers the first hurdle for anyone who does not want to assemble an environment from the command line. It does not make model files smaller or remove hardware limits and training time. Start with a small model and non-sensitive data. Check GPU compatibility, storage needs, and output quality before making it part of your routine.
7. OpenLogi
OpenLogi is a local-first alternative to Logitech Options+. It handles button mapping, DPI, and some scroll-wheel settings without requiring an account, and the project says it collects no telemetry. This is one of the most everyday tools in the ranking, but its README clearly says development is active and settings may change. Check the device support list first, keep a route back to the official software, and do not test it for the first time on your only work mouse.
8. OpenViking
OpenViking puts agent memory, knowledge retrieval, and skills in one context database. It may help teams that currently maintain several separate memory and knowledge tools because a unified interface can reduce synchronization work. The AGPL-3.0 license, data migration, and self-evolution mechanism all require review. If you only want one chat tool to remember your preferences, the architecture may be excessive.
9. omarchy
Omarchy is a Linux distribution led by DHH with a strong set of defaults, and it also appeared in the monthly ranking. Its selling point is not maximum choice. It configures the desktop, keyboard shortcuts, and common tools as a package. People who prefer keyboard-driven workflows and accept the author's trade-offs can try it on a spare machine or virtual environment. Anyone who needs corporate IT support, a specific VPN, or proprietary software should make a compatibility checklist first.
10. ai-memory
ai-memory handles long-term memory and handoffs between coding agents. Its README demonstrates leaving one agent midway through a task and having another take over. That targets a real problem in long-running work, but memory records may contain code, failed approaches, and sensitive context. Try it locally in one repository for a personal project. Before team use, define retention periods, exclusion rules, and who may read the data.
Top New Repos: When popularity arrives this fast, treat it as a watchlist
The top two new projects, dsh-routing-suite and dsh-anchored-standard, provide routing or presets for DeepSeek Harness. dsh-market, deepseek-harness-orange-book, and Deepseek-Harness-EAC belong to the same surrounding ecosystem. Five related projects entering the list in a very short period proves that attention is concentrated. It does not prove that the projects are independent, their settings are safe, or their benchmarks can be reproduced. If you do not already use Harness, start with the core project and its permission model instead of installing the entire extension set.
ip-as-logo-skill packages the visual rules for rounded mascot logos as an agent skill. It can help develop an initial direction, but it cannot replace trademark searches or a finished design process. zsvirt is an IaaS core written in Java. It is clearly infrastructure-oriented, so individual readers without a test environment or operational need can skip it for now.
Three other projects are closer to daily life. cumora puts people and agents in the same chat, board, and calendar, with options for a hosted agent or your own. The key questions are where the data goes and which tools the agent may operate. OpenBot gives each agent its own browser, files, and tools while logging actions. Its README still labels it alpha, and without a local identity provider it even treats requests as administrator requests, so isolated testing is the appropriate use today. northcinder requires a shopping agent to explain rankings, identify stores it did not search, and ask again before checkout. Its explain-then-approve approach is more useful than automatic ordering, although store coverage and ranking quality still need testing with your own shopping queries.
Community temperature: High engagement clustered around a few projects
The automated collection found many HN results, but keyword matches for public-apis, OpenLogi, omlx, and others initially included unrelated posts. Each result had to be checked against the project name and original discussion page. After that review, the main OpenLogi thread had 1,612 points and 425 comments. Supporters focused on replacing unstable official software, while doubts centered on device compatibility, marketing language, and whether a resident utility was needed at all. Needle 2's related thread had 535 points and 185 comments, and Unsloth Studio had 388 points and 82 comments. Several Omarchy discussions repeatedly debated whether it was a full distribution or an opinionated Arch and Hyprland configuration. These conversations reveal demand and friction, not product maturity.
On X, after filtering for engagement and account quality, the most useful diagram-design signal was the author's direct demonstration of a database schema output rather than generic reposts. OpenLogi discussion focused on mouse control without logins or tracking. OpenBot's CEO explicitly reminded readers that the product was still in alpha. MoneyPrinterTurbo attracted many high-engagement posts, but most relied on income claims or content-farm promotion and offered no concrete testing, so they were excluded. This week made the gap between popularity and credible experience unusually clear.
That distinction is easy to miss in a ranking. Stars help you discover candidates, and community discussion helps you find counterexamples and friction. Neither is a hands-on test. If you want to put an agent into real work or daily life, read our guide to evaluating long-running agent workflows, then test one project with a clear recovery path.
Monthly trend comparison
Only needle and Omarchy appeared in both this week's list and the monthly trend. Needle suggests that attention to small, on-device tool calling is still accumulating. Omarchy shows that an opinionated developer desktop is more than a one-day topic. The other projects only show short-term acceleration in this snapshot. Their presence does not establish a full month of sustained interest.
Three takeaways from this week
AI tools are moving from capability demos toward operational boundaries. ai-memory, OpenViking, semantica, and OpenBot address memory, context, provenance, and action logs from different angles. The questions now include what an agent remembers, why it acted, and how to stop it after an error.
Small models and local tools are reaching real devices. Needle 2, Unsloth, and OpenLogi face concrete limits such as RAM, GPUs, and mouse compatibility. For a typical user, reliably completing one task on their own device can matter more than answering every possible question.
Fast-growing ecosystems need an independence check. Several Harness presets, a market, and documentation projects appeared in the same week. That makes them worth watching, but their stars should not be added together as a proxy for market size. Check core dependencies, maintainers, licenses, and reversible installation methods before treating a burst of attention as a mature ecosystem.
If you want something you can inspect today with an easy way back, start with the diagram-design examples or OpenLogi's device support list. If you are planning an agent setup for a team, isolate one problem handled by ai-memory, semantica, or OpenBot. Do not introduce memory, permissions, and collaboration layers all at once. Testing one layer at a time makes it much easier to see where the time savings actually come from.
Was this article helpful?
