Aurora is a Go‑based self‑hosted server that turns the private ChatGPT web backend into an OpenAI‑compatible REST API, supporting chat, tool‑calling, image, audio, file upload, and token‑pool management.

Claude‑Code‑Proxy is a local command‑line service that translates Claude Code’s Anthropic API calls into requests for Codex, Kimi, Grok, OpenCode Go, or Cursor models, letting you use those providers while keeping Claude Code’s UI. It handles auth, model routing, streaming, and offers a terminal monitor for usage stats. Install via Homebrew or a script, run `claude-code-proxy serve`, and point Claude Code at `http://127.0.0.1:18765`. MIT‑licensed.

CodeWiki is an open‑source Python tool that uses large language models to automatically generate comprehensive, architecture‑aware documentation (markdown + Mermaid diagrams) for whole codebases across ten languages. It hierarchically decomposes a repository, runs a recursive multi‑agent LLM pipeline, and outputs ready‑to‑publish docs, with support for many LLM providers and a rich CLI.

Puppetmaster is a Python‑based supervisor that orchestrates existing LLM coding assistants (Cursor, Claude‑Code, Codex, Gemini Antigravity, Hermes, etc.) by routing tasks to independent workers, persisting each typed result in SQLite, and providing a stitched summary and dashboard. It adds durable, auditable state, cost‑aware routing, and parallel fan‑out, enabling reliable repository audits, refactors, and implementations. The project is actively maintained (v1.27.8 beta), MIT‑licensed, and includes benchmark evidence of cost savings and higher pass rates.

Proma is an open‑source, cross‑platform desktop AI agent that combines chat, planning, project‑based memory, and embedded tools (browser, terminal, calendar, Obsidian‑style markdown editor). It supports parallel sub‑sessions, skill bundles, and remote connectors (Slack, WeChat, etc.). Installable via binaries for macOS, Windows, and Linux; the open‑source edition requires user‑provided LLM APIs, while a commercial edition adds hosted model access, team management, and extra services. Licensed under AGPL‑3.0 with optional commercial licensing.

A collection of 30 Jupyter notebooks that teach and demonstrate the main memory patterns for LLM‑based agents—from short‑term buffers to long‑term vector stores, cognitive‑architecture layers, and production‑grade frameworks (Mem0, Letta, Zep, Graphiti). Includes decision tree, comparison matrix, and links to a free video course and a printed guide.

Alexandria is a locally‑run web app that converts a text file (txt/md/epub) into a complete audiobook. It uses an LLM to annotate the book with speaker tags and TTS instructions, then synthesises each line with the built‑in Qwen‑3 TTS engine. Users can pick from nine preset voices, clone a voice from a short sample, fine‑tune voices with LoRA, or generate new voices from textual descriptions. The UI lets you edit the script, batch‑render audio (GPU‑accelerated), and export as MP3, M4B, or per‑speaker WAV tracks for Audacity.

Investment News is a Python‑only, locally‑run dashboard that fetches >100 global news feeds, uses a user‑provided LLM (Claude subscription or any OpenAI‑compatible API) to summarise each of 12 A‑share sectors into 3‑5 Chinese bullet points per day, translates headlines, and displays everything in a single browser page. No third‑party packages, no cloud service, and all data stays on the user’s machine.

nurb is a locally‑run Python/macOS app that lets you describe a 3‑D part in plain language, has your existing LLM write real B‑Rep CAD code, checks the model against 13 print‑ability rules, shows live updates in a three.js viewer, and exports print‑ready 3MF (or STL/STEP/GLB) with slicer settings embedded.

Firefox‑Reverse is a custom Firefox build that embeds an AI‑driven reverse‑engineering agent. The agent can automatically capture, trace, and reproduce encrypted request parameters (signatures, tokens, risk fingerprints) generated by obfuscated JavaScript, JSVMP, or WASM. It provides isolated fingerprint environments, 68 built‑in tools, model‑agnostic LLM support, and outputs ready‑to‑run Node.js/Python scripts (or pure‑JS code). Pre‑built installers are available for Windows, macOS, and Linux.

A collection of 238 ready‑to‑use “skills” (instruction + code templates) that let AI coding assistants automate common construction‑company tasks such as cost‑database search, BIM‑to‑Excel conversion, report generation, and AI‑driven analytics. Organized into six categories—from production tools to experimental AI/ML pipelines—each skill is a `SKILL.md` file that the assistant reads, generates code for, and the user then adapts and runs. Requires Python 3.9+, an AI coding assistant, and the company’s own data (Excel, IFC, PDFs, photos).

Cookiecutter Data Science (CCDS) is a CLI tool that creates a standardized, best‑practice project skeleton for data‑science work, including folders for raw/processed data, notebooks, a Python package with training/prediction scripts, docs, and a Makefile. Install via pipx, run `ccds`, answer prompts, and you get a ready‑to‑use, reproducible project layout.

An open‑source Python framework for training, evaluating, and running inference with many state‑of‑the‑art music source‑separation models. It provides config‑driven training (single‑GPU or multi‑GPU), LoRA fine‑tuning, a CLI, a small GUI, and scripts for ensembling and converting models to ONNX/TensorRT.

A system-wide speech-to-text and AI text processing tool for macOS that supports local on-device models or cloud APIs for transcription and transformation.

projectmem is a Python‑based, local‑first memory layer for AI coding assistants. It records structured events (issues, attempts, fixes, decisions) in a `.projectmem` folder, provides pre‑commit warnings to avoid repeating failures, and exposes the data via a native MCP server that Claude, Cursor, Codex and similar tools can query. With a single command (`pjm init`) it installs git hooks, starts a cross‑platform file watcher, and creates dashboards for visualizing the project’s debugging history and ROI. The tool aims to reduce token usage, prevent repeated bugs, and keep AI agents “experienced” across sessions.

Anthropic’s *k12-teacher-skills* repo provides four Claude‑agent plugins for lesson‑plan creation, differentiation, prep, and formative checks, all built to use the Learning Commons Knowledge Graph for standards alignment. The skills are pre‑installed for Claude for Teachers users and can be added manually elsewhere; the repo also includes an evaluation framework for testing and adapting the skills.

A prompt‑template + tiny CLI that structures AI‑assisted coding into five clear steps (research, PRD, tech design, agent‑file setup, build). Works with Claude, Gemini, ChatGPT, Cursor, VS Code, etc., and generates files like AGENTS.md and MEMORY.md to keep the AI on track. Ideal for solo devs or small teams wanting a repeatable, safety‑aware workflow for rapid MVP creation.

A Python research project that compares four modern time‑series models (PatchTST, TFT, N‑HiTS, CatBoost) for hour‑by‑hour electricity‑load forecasting using real Russian utility data. Includes data preprocessing, training pipelines, statistical analysis, and visualisation, and is released under Apache 2.0.

Headlong is a Bash‑based micro‑harness that runs a continuously thinking LLM‑backed agent. The core `shellm` engine lets the model write Bash commands, executes them, and feeds the output back as context. Agents persistently generate thoughts, can be accessed via Slack, Telegram, or a web dashboard, and store their entire history as a JSON‑L DAG with tiered summarisation. Installation is a one‑liner that sets up Docker sandboxing (or host mode) and a unified `llm` CLI for Anthropic, OpenAI, Gemini, OpenRouter, or any local OpenAI‑compatible server (Ollama, llama.cpp, vLLM, etc.). The framework emphasizes tiny, composable Bash tools, self‑improvement via fork‑test‑merge, and safety via sandboxing and spend‑capped API keys.

Open Island is a free, native macOS app that lives in the notch or menu bar and provides a live dashboard for AI coding agents (Claude Code, Codex, Cursor, Gemini‑CLI, etc.). It shows session status, handles permission prompts, displays usage stats, and lets you jump back to the exact terminal or IDE pane with a click. Integration is hook‑based: agents emit JSON events, a tiny CLI forwards them via Unix socket to the app, which updates its UI. Supports 13 agents and 15+ terminals/IDEs, runs locally with no telemetry, and is distributed as a signed DMG or Homebrew cask.

A plugin that equips AI coding assistants (Claude Code, Codex, Cursor, VS Code Copilot, etc.) with Shopify‑specific capabilities: searchable docs, API‑schema lookup, GraphQL/Liquid validation, and store‑management commands, all usable directly from the editor or terminal.

OpenTabs is a CLI‑plus‑Chrome‑extension toolkit that lets LLM agents call real web APIs (Slack, Discord, GitHub, etc.) using the user’s existing browser session, with a library of ready‑made plugins, local‑only execution, and a permission‑based security model.

Apify Agent Skills is a set of plug‑in style “skills” that let LLM agents (Claude, Cursor, Gemini, etc.) invoke Apify’s web‑scraping/automation platform. Five core skills cover universal scraping via 130+ curated Actors, creating new Actors, turning existing code into Actors, auto‑generating output schemas, and building Apify integrations. Install with a single `npx skills add …` command, then prompt the agent to scrape sites, generate leads, do market research, etc. Requires an Apify account, the Apify CLI, and Node 20+. Licensed Apache‑2.0.

PyGPT is an open‑source desktop AI assistant (Linux, Windows, macOS) that lets users chat with a wide range of LLMs—both cloud‑based (OpenAI, Gemini, Claude, etc.) and locally hosted via Ollama. It supports multimodal interactions (text, files, images, video, speech), RAG through LlamaIndex, multi‑agent orchestration, code execution, and extensible plugins, all wrapped in a Qt‑based GUI.

zerostack is a Rust‑written terminal AI coding assistant that talks to many LLM providers, offers multiple built‑in prompts (code, plan, review, debug, etc.), enforces a granular permission system, supports session persistence, optional sandboxing, and extensible features like advisor models, hooks, and status‑signal sockets—all while staying lightweight (≈ 16 MiB RAM, 26 MB binary).

Codex‑Manager is a Rust/Tauri desktop and service app that lets you pool, manage, and route OpenAI‑style API keys (Codex, Gemini, Claude, etc.). It provides UI‑driven account import/export, quota tracking, automated reset reminders, a local OpenAI‑compatible gateway, model routing, and a plugin/skill ecosystem.

OpenSquilla is a token‑efficient, micro‑kernel AI‑agent framework that routes each turn to the cheapest LLM, provides on‑device memory, web search, and a sandboxed tool layer, and runs on Windows/macOS/Linux via desktop installers, a quick‑install Python wheel, or source builds.

Mirage is a library that gives AI agents a unified, Unix‑like terminal backed by a virtual filesystem. All data sources (S3, Google Drive, Slack, Redis, email, GitHub, etc.) are mounted under one root, and the agent can use familiar CLI tools and sandboxed runtimes (Python, JavaScript) to read/write data, run scripts, and invoke virtual commands. Fine‑grained policies control what the agent can see or do, improving security and token efficiency.

Long‑Horizon Terminal‑Bench (LHTB) is a 46‑task benchmark that measures how well LLM agents can keep working inside a Docker‑based terminal for hundreds of steps. It ships a patched version of the open‑source Harbor harness that adds a “continue‑until‑timeout” loop and sandboxed verifier isolation, preventing agents from cheating. Results show current frontier models solve only a minority of tasks, making LHTB a hard, frontier‑level evaluation suite for autonomous AI agents.

genjutsu is a Claude‑AI plugin that lets you ask an LLM to add or redesign motion‑heavy UI elements. Two main skills – `/genjutsu:cast` for small, targeted animations and `/genjutsu:paint` for full‑scale redesigns – automatically detect your stack (React/Vue/Svelte, Compose, SwiftUI), propose an interaction thesis, generate the code (GSAP, Framer Motion, Compose, SwiftUI, etc.), and run a quick audit for accessibility and performance. Install via a single ZIP bundle or individual skill ZIPs on Claude AI, Claude Code, or Cowork.

A real‑time monitoring dashboard for Claude Code and OpenAI Codex agents, built with Node/Express, React, SQLite, and Electron, offering live visualisations, analytics, and desktop/IDE integrations.

Live Avatar is a research‑grade system that turns spoken audio (plus an optional reference image or text prompt) into a continuous, real‑time video of a talking avatar. Built on a 14 B diffusion backbone (WanS2V‑14B) with a LoRA fine‑tune, the authors achieve ~45 FPS streaming on multi‑GPU H800 hardware and can generate videos longer than 10 000 seconds via block‑wise autoregressive processing. The repo provides Conda‑based installation, multi‑GPU and single‑GPU inference scripts, optional FP8 quantisation and JIT compilation for speed, and a Gradio UI. Model weights are hosted on Hugging Face under Apache 2.0.

A fork‑ready framework that uses Anthropic Claude Code to automate academic tasks (slides, papers, data analysis, etc.) via specialised agents, a ten‑gate quality suite, and extensible skills; includes setup scripts, pre‑commit hooks, and detailed documentation.

A Python package for audio stem separation that uses UVR models to split audio files into vocals, instrumentals, and other specific instrument tracks.

TalkingHead is an open‑source JavaScript library for browser‑based 3‑D avatars that can lip‑sync, show facial expressions, and be driven by AI services (Google TTS, OpenAI, Whisper, etc.). It works with Mixamo‑rigged GLB models, supports multiple languages, and offers plug‑in modules for TTS, audio‑driven visemes, and advanced gestures. The README provides demos, integration steps, and a large set of configurable options, making it suitable for research, product demos, or hobby projects involving embodied AI.

A Claude Code plugin that transforms a conversation into a full album production pipeline, coordinating research, lyrics, and audio mastering for AI-generated music.

An open-source, self-hosted companion app for AI voice recorders that syncs recordings and transcribes them using user-chosen AI providers.

A Navidrome plugin that integrates AI-powered song and artist similarity features from AudioMuse-AI into the Navidrome music server frontend.

Qiushi‑Skill is a plug‑in library that adds nine philosophically‑derived reasoning methods (e.g., contradiction analysis, investigation‑first, mass‑line) to AI agents such as Claude Code, Cursor, and Codex. Installed via `npx qiushi-skill` or the Claude Code marketplace, it injects a lightweight kernel that enforces the principle “seek truth from facts” and routes tasks to the appropriate skill. Each skill provides trigger rules, step‑by‑step actions, output templates, hard constraints, and hand‑off specs, plus reference texts and optional sub‑agents for investigation or self‑criticism. The project aims to make agents more evidence‑driven, self‑critical, and strategically focused.

A Codex plugin that enables an AI agent to draw scientific figures live inside the desktop draw.io editor by calling draw.io’s graph API via a local MCP server. It provides tools for launching draw.io, adding shapes/edges, inspecting the canvas, saving the diagram, and exporting to PNG/SVG/PDF. The repo is legacy; active development has moved to the `scientific-illustrator` project.

Hunter Community Edition is an open‑source, Docker‑compose‑deployed financial AI agent platform. With a single LLM API key it provides real‑time market data (32+ sources), 23 pre‑built analysis SKILLs, a local “investment thesis” memory, and plug‑and‑play model compatibility (DeepSeek, Claude, Qwen, etc.). Everything runs on the user’s machine—data, chat history, and thesis stay on local disk—while remaining extensible via custom Markdown SKILLs or user‑provided MCP data services.

Reversa is a Node‑based framework that installs into a legacy codebase and orchestrates specialized AI agents (Claude Code, Codex, etc.) to extract hidden business rules, architecture, data models, and UI tokens, turning them into traceable, machine‑readable specifications. The installer (`npx reversa install`) creates only new files (`.agents/`, `.reversa/`, `_reversa_sdd/`) and never modifies existing code. Core commands include `/reversa` (full Discovery pipeline), `/reversa‑autonomous` (unattended run), `/reversa‑new` (green‑field spec generation), `/reversa‑forward` (spec‑to‑code), `/reversa‑migrate` (legacy‑to‑modern migration), and `/reversa‑docs` (HTML mini‑site). Designed for safe, human‑in‑the‑loop reverse engineering, it enables AI agents to safely evolve, refactor, or rebuild legacy systems.

A local desktop application that uses offline speech recognition and AI to help users improve their oral expression by identifying filler words and vague language.

MacParakeet is a native macOS app (and CLI) for offline, local‑first voice dictation, audio/video file transcription, and meeting recording on Apple Silicon. It runs several Core ML‑based speech‑to‑text models (Parakeet v3, v2, Unified) and optional local Nemotron, Cohere, or WhisperKit engines, keeping data private. Optional AI providers can be hooked for summaries or text rewriting. The stable DMG (0.8.7) is downloadable directly or via Homebrew, and the CLI enables automation of all features.

Observer AI is an open‑source desktop/web app that lets you create tiny, locally‑run AI agents (micro‑agents). Each agent can watch your screen, webcam, microphone, clipboard, etc., and react with notifications, messages, recordings, or system actions using a set of built‑in JavaScript tools. It runs on local LLMs (Gemma, llama.cpp, or any OpenAI‑compatible endpoint) so all data stays private.

An audio fingerprinting library that generates compact identifiers to detect near-identical audio files and streams for high-performance searching.

DeterminFlow is an open‑source, AGPL‑licensed framework for building, running, and monitoring production‑grade AI workflows. It breaks complex pipelines into isolated LLM‑Agent nodes, scripts, API calls and human approvals, providing versioning, checkpoint‑based recovery, per‑node token accounting, and a visual editor. The tool is aimed at content generation, data‑centric AI pipelines, and enterprise automation, and can be installed via desktop binaries, run from source, or deployed with Docker.

Zeraix is an open‑source desktop app that lets you run local LLMs (GGUF models) on macOS or Windows, chat with them, and have them read/edit files, run terminal commands, and manage multi‑step “Skills”. It includes a Rust agent runtime, a media library, plug‑in support, optional QEMU sandboxing, and can also connect to cloud models. Installers are provided; the project is licensed Apache‑2.0 and backed by a Discord community.