This Week's Top 14 GitHub Open Source Projects: AI Agents, DevTools & Voice Studios

01 Scientific Agent Skills

Scientific Agent Skills collects over 160 research skills and documents usage for more than 100 scientific databases. Skills span bioinformatics, drug discovery, medical research, materials science, geospatial analysis, and academic writing. Each skill specifies which tools an agent should call, which processes to follow, and which results to verify. The core value is codifying research workflows into reusable procedural knowledge so models avoid dead ends and complex pipelines become reproducible. The author cautions that installing the resource does not grant AI scientific judgment; researchers must still validate data sources, statistical methods, and experimental conclusions. Repository:

https://github.com/K-Dense-AI/scientific-agent-skills

02 vphone-cli: Virtual iPhone on Mac

vphone-cli runs a virtual iPhone inside a Mac, controllable via automation interfaces for tapping, swiping, and screenshots. It targets developers who need to verify flows like app launch → login → input → result without repetitive manual testing. Requirements are steep: M-series Mac, macOS 15+, Xcode, and adjustment of macOS security settings. Repository:

https://github.com/Lakr233/vphone-cli

03 OpenSEO: Open-Source SEO Platform

OpenSEO provides keyword research, rank tracking, competitor analysis, backlink analysis, and site audits. It offers a web UI, an MCP server, and a set of SEO skills so agents can read SEO data and perform keyword clustering, competitive landscape analysis, and backlink opportunity aggregation. Deployment supports Docker and Cloudflare. The code is free, but SEO data incurs cost via DataForSEO API (pay-per-call). The local Docker default runs in single-user mode without authentication and should not be exposed publicly without configuration. The project saves software subscription fees and workflow lock-in; data procurement costs remain. Repository:

https://github.com/every-app/open-seo

04 Heretic: Reducing Model Refusal Behavior

Heretic modifies AI models to weaken refusal responses (e.g., "I'm sorry, I can't answer that") while preserving original capabilities. The author frames it as adjusting the "which questions to refuse" layer of a trained assistant. The tool aims for automatic removal of model censorship. Repository:

https://github.com/p-e-w/heretic

05 Chinese Patent Skill

This skill structures patent drafting into a workflow: invention point mining, prior-art search, disclosure drafting, and revision management. Users feed design docs, code, Word, or PPT to the agent; it scans materials, proposes novel points, prioritizes searches via the CNIPA patent publication site, and generates Markdown and Word documents with system block diagrams and flowcharts. Revisions are saved as new versions with change logs for collaboration between R&D staff and patent attorneys. Repository:

https://github.com/handsomestWei/patent-disclosure-skill

06 VoiceStudio: Local Voice Studio

VoiceStudio is a cross-platform (macOS, Windows, Linux) open-source voice suite supporting voice cloning, voice design, video dubbing, dictation, transcription, and audiobook creation. Users create voice samples from short audio, batch-generate speech, and process video through transcription, translation, speaker diarization, and re-dubbing. It exposes a local API, OpenAI-compatible audio endpoints, and an MCP server for integration into scripts or agent workflows. Engines run on CUDA, Apple Silicon, or CPU. Repository:

https://github.com/debpalash/VoiceStudio

07 OpenClaude: Terminal Coding Agent with Model Switching

OpenClaude is an open-source terminal coding agent (32k+ stars) that lets users switch between cloud APIs, local models, and multiple models within a single workflow. It retains standard coding-agent capabilities: file read/write, search, shell, MCP, skills, and sub-tasks, while returning model selection control to the user. Repository:

https://github.com/Gitlawb/openclaude

08 TimesFM 3.0: Google's Multivariate Time-Series Foundation Model

TimesFM, from Google Research, now supports native multivariate forecasting. Typical domains: inventory, sales, energy, traffic, device metrics. Traditional time-series projects require per-dataset model selection, training, and tuning. TimesFM emphasizes zero-shot prediction — forecasting without seeing the target training set. Version 3.0 handles multiple correlated series simultaneously and incorporates historical covariates and future known inputs (holidays, prices, weather forecasts). The release pushed the project onto the GitHub Trending list. Repository:

https://github.com/google-research/timesfm

09 MiniMind: Compressed LLM for Learning Training Internals

MiniMind compresses a large language model's architecture and training pipeline into a scale (~64M parameters) that an individual developer can understand, reproduce, and modify (58k+ stars). The repo covers data cleaning, pre-training, supervised fine-tuning, LoRA, preference optimization, reinforcement learning, tool use, and model distillation. Core algorithms are implemented in raw PyTorch so every step is visible without high-level framework abstraction. Repository:

https://github.com/jingyaogong/minimind

10 Awesome MCP Servers

A curated list of MCP servers callable by Codex, Claude Code, and similar tools (94k+ stars). Entries are categorized by browser automation, databases, dev tools, file systems, search, finance, multimedia, etc. Each entry notes primary language, local vs. cloud execution, supported OS, and official status. The author notes that browsing the categorized directory is faster than web search when first wiring tools to an agent. The list mixes production-grade and experimental implementations. Repository:

https://github.com/punkpeye/awesome-mcp-servers

11 Screenshot to Code

Converts page screenshots, design mockups, and screen recordings into runnable frontend code (77k+ stars). Supports HTML, CSS, React, Vue, and multiple styling solutions. Generated output can be previewed in-browser and iteratively refined. Video mode captures interaction sequences, restoring some dynamic behavior that static screenshots miss. Repository:

https://github.com/abi/screenshot-to-code

12 ipatool: CLI for App Store Search & Download

ipatool runs on Windows, Linux, and macOS to log into the App Store, search apps, and download encrypted IPA packages. It lists historical versions and supports downloading a specific version via external version ID. Useful for app archiving, compatibility testing, and security research. Output supports text and JSON; non-interactive mode enables script integration. macOS install via brew install ipatool. Requires an Apple ID with App Store access. Repository:

https://github.com/majd/ipatool

13 OpenWhispr: Cross-Platform Voice Input for Desktop Workflows

OpenWhispr writes spoken words at the current cursor position and handles meeting transcription, note organization, and voice-agent operations. Users can choose local Whisper models (audio stays on-device) or cloud APIs for speed. Meeting mode detects Zoom, Teams, FaceTime calls, provides speaker diarization, note search, and follow-up AI actions. An MCP server exposes transcripts and notes to other agents. Repository:

https://github.com/OpenWhispr/openwhispr

14 Archify: AI-Generated Architecture Diagrams

Archify turns a system description or a code repository into an interactive system diagram (48k+ stars, rapid weekly growth). The pipeline constrains the agent to emit typed structured data; a validator checks structure, layout, and composition rules before rendering HTML and SVG. Nodes, edges, and paths derive solely from the validated data — the renderer adds no extra dependencies. Validation does not guarantee the diagram matches the real system. Supported diagram types: architecture, workflow, sequence, data-flow, lifecycle. Output is a self-contained HTML file with node search, upstream/downstream path tracing, and export to PNG, SVG, and video. Repository:

https://github.com/tt-a1i/archify