GitHub AI Weekly · Issue 04 (Sep 5, 2026) · AI Beyond

Top 10 projects

1. 01

   tt-a1i/archify

   AI codingJavaScript

   This week+24,227
   

   An agent skill that compiles typed intermediate data into verifiable architecture, workflow, sequence, data-flow and lifecycle diagrams, delivered as self-contained HTML.

   Stars

   48,036

   Forks

   3,100

   Open issues

   112

   Last commit

   2026-09-04

   It is charting for a second consecutive week, and this one is not momentum alone: v2.16.0 shipped on Aug 30 with a constraint-driven compiler, reproducible archives, a localized viewer and a RAG architecture example, followed by fixes for license provenance, preview-watcher fallback and CLI argument validation.

   Developers already asking Codex, Claude Code or OpenCode to read repositories can try it on one real system when they need the structure turned into a reviewable artifact.

   The diagram captures an agent's interpretation of code and documentation, not ground truth; using it in review still requires source checks, and 112 open issues show that a fast-growing project has plenty of edges left.

2. 02

   bilawalsidhu/gods-eye-view

   AI agentsJavaScript

   This week+10,485
   

   A browser-based 3D spatial-intelligence interface that layers public aircraft, vessel, satellite, earthquake, traffic and camera feeds onto a globe, with control through a realtime voice agent.

   Stars

   17,493

   Forks

   3,528

   Open issues

   124

   Last commit

   2026-09-04

   The project opened its source and released v0.1.0 and v0.1.1 this week, turning the showcase videos into something people could launch keyless through Pinokio; the patch then repaired macOS startup, dropped fire batches and response-size handling for live feeds.

   People building maps, public-data tools, OSINT visualizations or spatial-interaction prototypes can pull apart its data layer, Cesium scene and voice tool calls.

   Do not mistake the cinematic surface for uniform data precision: keyless traffic is simulated, camera poses and rocket trajectories are coarse estimates, and voice requires a metered OpenAI API; GitHub currently identifies no standard license, so reuse also needs a rights check.

   Deep dives

   The visual makes this easy to misread as a movie-control-room demo. The more useful part to study is the orchestration underneath it. Aircraft, vessels, satellites, earthquakes, public cameras and bike stations come from different open feeds with different refresh rates, coordinate quality and authentication. God's Eye View pulls those signals into one spatial interface, then lets a user move from an object on the globe to its current state.

   AI is only one layer, but it has a concrete job. The voice agent is wired to 28 action tools for switching layers, finding places, selecting nearby contacts and framing the camera; at street level it can also inspect the current view for visible signs. Without an OpenAI key the globe still works, while voice and the AI HUD summary switch off. That graceful degradation is healthier than binding the entire application to one model endpoint.

   The more the screen resembles an intelligence system, the more its source notes matter. The repository says plainly that some traffic is simulated along real roads, camera orientation and rocket trajectories are estimates, and free and paid feeds carry different quotas. It is a compelling interactive laboratory for public data. Using a screenshot as an operational claim goes beyond what the project can guarantee.

3. 03

   THU-MAIC/OpenMAIC

   AI agentsTypeScript

   This week+10,023
   

   An open multi-agent classroom builder that starts from a prompt or uploaded documents, audio, video and web material, then generates and revises lessons, quizzes, interactives, images and narration.

   Stars

   31,497

   Forks

   5,183

   Open issues

   217

   Last commit

   2026-09-04

   The Aug 27 v1.0.0 release replaced one-shot generation with a conversational course workbench whose runs can be paused and resumed, and that launch carried into this week's board; recent commits then added concise session titles, Exa search and double-click canvas text while bounding media requests, skill retries and archive expansion.

   Teachers, training teams and education-product builders who need existing material turned into a playable course prototype can validate the structure in the hosted demo before deciding whether to self-host.

   Facts, answer keys and media rights still need a page-by-page human review; self-hosting requires at least one model provider or a local Lemonade setup, while durable storage, media extraction and video export add Postgres, FFmpeg and container work, and the example persistence authentication is explicitly not production-ready.

   Deep dives

   OpenMAIC used to look like “enter a topic and wait for a classroom.” Version 1.0 turns it into a production desk you can negotiate with. The agent plans the curriculum and builds page by page; you can upload documents, audio and video as source material, cancel or resume a run, and ask it to rebuild one part. That is not the flashiest change, but it determines whether the tool can leave demo territory and enter real lesson preparation.

   The difficult part of a long course is not producing one attractive page. It is keeping page fifteen attached to the same learning objective and recovering without restarting the whole job. This week's commits expose exactly those engineering problems: media requests need unique IDs, scene-generation retries need a ceiling, bottom table rows cannot disappear in playback, and failure causes must survive. Those are closer to production needs than another animation preset.

   It is ready to try, not ready to turn “course generated” into “course complete.” Material extraction can miss context, a model can make a smooth explanation wrong, and every quiz deserves its own check. A shared deployment also has to replace the sample authentication intended only for localhost or a trusted network. The safest role today is first-draft structure and an interactive prototype, with a teacher retaining publication control.

4. 04

   K-Dense-AI/scientific-agent-skills

   AI agentsPython

   This week+7,370
   

   A skills collection for research agents spanning bioinformatics, chemistry, medicine, drug discovery and scientific visualization, with access to more than 100 databases; the repository description says 165 skills while the README currently lists 163.

   Stars

   42,632

   Forks

   3,899

   Open issues

   31

   Last commit

   2026-09-02

   This week's concrete additions are paper arXiv:2609.00065 and v2.66.0, alongside a shift from Claude-specific wording to the open Agent Skills standard; the release itself is mostly documentation, citation and scanner-report work rather than a sudden batch of new research capabilities.

   Researchers already doing analysis, literature retrieval or computational biology in Codex, Claude Code, Cursor or Gemini CLI are better served installing a small topical subset than importing the entire collection.

   The repository itself warns against installing everything at once: a skill can execute code, install packages, use the network and modify files, while community submissions cannot be exhaustively reviewed; individual skills also bring their own databases, credentials and Python dependencies, and their output does not replace research validation or clinical judgment.

5. 05

   debpalash/VoiceStudio

   MultimodalPython

   This week+3,902
   

   A local-first voice workbench for dubbing, voice cloning and design, transcription, dictation and audiobook workflows; the repository lists 16 TTS engines, 11 ASR engines and a 646-language catalog.

   Stars

   17,928

   Forks

   2,354

   Open issues

   24

   Last commit

   2026-09-04

   Momentum from v0.5.1 was still running when this week's main branch added watch-folder ingestion, realtime dubbing preview and a simpler voice-design panel, then fixed non-English Windows install paths, stale backend processes and several localized interfaces.

   Podcast, audiobook and video teams that want recordings and models to stay on their own machines can choose an engine around their language and hardware rather than a single cloud plan.

   The 646-language figure is the union of the catalog, not equal quality in every engine; the default local workflow asks for at least 8 GB of RAM and 10 GB of disk, accelerated work starts around 4 GB of VRAM with 8 GB preferred, the application is AGPL-3.0, and downloaded models keep their own terms.

6. 06

   jingyaogong/minimind

   InferencePython

   This week+3,122
   

   An educational, from-scratch small-language-model repository covering data cleaning, pretraining, SFT, LoRA, DPO, PPO, GRPO, tool use, agentic RL, distillation and serving, with a roughly 64M-parameter dense model on the main path.

   Stars

   58,527

   Forks

   7,605

   Open issues

   60

   Last commit

   2026-09-04

   The repository kept exposing the experiment surface this week: random seeds and MoE expert configuration reached the CLI, benchmark tables gained resolution notes, and a confusing expert-argument path was then removed. The attraction is not a new frontier model but a training stack people can follow end to end.

   Students and engineers who want to run the entire language-model training, alignment and serving path themselves can learn more cheaply on 64M parameters than by starting inside a multi-billion-parameter codebase.

   The README's two-hour claim refers specifically to one SFT epoch on a single RTX 3090, not complete pretraining plus alignment in two hours; a 64M model is useful for teaching and testing code paths, not as a proxy for production-model quality.

7. 07

   every-app/open-seo

   AI agentsTypeScript

   This week+2,941
   

   A self-hostable SEO workbench for keyword research, rankings, competitors, backlinks, audits and AI visibility, exposing its data to Claude Code, OpenClaw and other agents through MCP and Agent Skills.

   Stars

   17,017

   Forks

   2,117

   Open issues

   150

   Last commit

   2026-09-03

   Version 0.1.7 on Sep 2 added deeper SERP loading on demand and repaired Search Console Discover and Google News reports in MCP, crashes during tool-heavy answers, and blank screens in older browsers; the two commits after the release are blog content only.

   Small teams already willing to pay for underlying SEO data, and wanting an agent to query it and run fixed research workflows, can self-host it through Docker or Cloudflare.

   Open source does not make the data free: self-hosting still requires a DataForSEO key and usage charges, while the hosted service adds a margin; 150 open issues and several external data dependencies make this an adaptable young product, not a zero-cost replica of Semrush's coverage.

8. 08

   abi/screenshot-to-code

   MultimodalPython

   This week+2,412
   

   A tool that converts screenshots, designs and screen recordings into HTML, Tailwind, React, Vue, Bootstrap or Ionic prototypes, with an optional browser preview for visual checking.

   Stars

   77,685

   Forks

   9,454

   Open issues

   141

   Last commit

   2026-08-14

   The repository has had no new commit since Aug 14, so this week's 2,412 stars are renewed circulation around an established project, not evidence of a feature release.

   People who need a runnable front-end skeleton from a competitor screenshot or static design get the clearest time saving by treating the output as a first pass and then taking over components and interactions.

   Local use needs at least one OpenAI, Anthropic or Gemini key, with Gemini and Replicate separately recommended for asset extraction and image work; a screenshot cannot reveal business logic, responsive rules or accessibility semantics, every output needs review, and the maintainers do not recommend local Ollama models for final quality.

9. 09

   google-research/timesfm

   InferencePython

   This week+2,324
   

   A pretrained time-series foundation model from Google Research for zero-shot forecasting; version 3.0 natively handles univariate and multivariate series plus past-only and past-and-future dynamic covariates.

   Stars

   31,041

   Forks

   2,956

   Open issues

   236

   Last commit

   2026-09-04

   TimesFM 3.0 entered this week's board after its Aug 28 release; commits this week did not add another headline feature, but fixed KV-cache indexing in batched decoding, covariate-window alignment after context truncation, and rejection of unsupported SwiGLU configuration.

   Data teams with sales, traffic, capacity or sensor series can use it as a zero-shot baseline before comparing it against statistical models and task-specific training.

   The code is Apache-2.0, but the TimesFM 3.0 pretrained weights currently use a separate non-commercial license that explicitly forbids commercial and production use; the repository also says this open version is not an officially supported Google product, and the first-week inference fixes make version pinning and regression tests essential.

   Deep dives

   Last issue had no model repository in the top ten. TimesFM and MiniMind bring that slot back from opposite directions. TimesFM 3.0 is more than a checkpoint refresh: it makes multivariate series and dynamic covariates native inputs, so one model can consider several related signals and future-known information such as holidays or campaign schedules. For sales and capacity forecasting, that is much more useful than extending one historical column.

   The first week after release also shows that 3.0 is still settling. Incorrect KV-cache indexing during batched decoding and misaligned covariate windows after context truncation are the kind of bugs that can produce a plausible but wrong result without crashing. Existing 2.x users should not judge an upgrade by sample plots alone; they need to replay their own backtests case by case.

   The license, not the chart position, decides whether this can ship. Repository code is Apache-2.0 and weights through 2.5 remain Apache-2.0, but the 3.0 weights are currently restricted to non-commercial, non-production use. Research baselines can start immediately. A company either stays on an older permitted checkpoint or confirms terms through Google's supported product routes; open source code does not make the newest model commercially open.

10. 10

    handsomestWei/patent-disclosure-skill

    AI agentsPython

    This week+1,846
    

    An agent skill for Chinese patent work, assisting with invention mining, disclosure drafts for invention, utility-model and design patents, plain-language reading of published patents, and office-action response drafts.

    Stars

    7,368

    Forks

    814

    Open issues

    7

    Last commit

    2026-09-04

    This week it split different jobs into separately installable skill packages, restored CNIPA classification fixtures, preserved source formulas, and moved Word generation for office-action responses behind explicit user confirmation — all changes that tighten the document-delivery path.

    Chinese R&D teams that need to structure project material into a first disclosure draft, read published patents in batches or prepare office-action material can use it as a workflow assistant.

    It does not replace a patent attorney or legal advice, and prior-art scope, claims and response strategy all require professional review; the workflow also depends on Python, Playwright, public CNIPA pages and local document conversion, so upstream page changes can require maintenance.