An LLM‑orchestrated CLI tool that turns a high‑level request into a spec, spawns Claude‑based design agents and Codex‑based code‑generation agents, runs isolated builder worktrees, validates with frozen checks, and finally emits a single PR or local finish record—no human approval gates required.

DreamCoder is an open‑source native desktop GUI (built with Tauri 2, React 18, and Bun) that wraps the Claude Code LLM. It provides multi‑model support, visual session management, integrated terminal, and transparent tool‑calling, with a privacy‑first design that stores API keys locally.

A demo that combines OpenAI GPT, Whisper, Eleven Labs TTS, and Rhubarb Lip‑Sync to create a talking 3‑D avatar with facial expressions and body animations. Install dependencies, set API keys, run `yarn dev`, and interact via text or voice.

An autonomous 24/7 AI radio station that uses agents to generate music, write hosted talk breaks, and manage live streaming via Icecast.

A Japanese full-duplex spoken dialogue system that enables natural, real-time voice interactions with overlapping speech and backchanneling.

dsh‑memory‑evolve is a DSH plugin that gives the AI persistent cross‑session memory, todo/project logs, Git‑based memory sync, multi‑agent orchestration (internal sub‑sessions and external AI providers), a searchable infinite canvas for files, optional reviewer sessions, and mobile‑friendly UI with notifications. It lets the AI act as a long‑term partner and team coordinator while keeping the user in control of all writes.

A high-speed, accurate speaker diarization pipeline that identifies who spoke when in audio recordings, capable of processing one hour of audio in seconds.

TapQ is a Swift‑based macOS runtime that connects coding agents (Claude Code, Codex, Cursor, OpenCode) to your AirPods. It reads agent prompts aloud, lets you answer with double‑nod/shake gestures, stem swipes, or voice (optionally via OpenAI’s realtime API), and falls back to the on‑screen UI if it can’t interpret you. All motion processing is on‑device; audio is sent to OpenAI only during a short response window. Supports macOS 14+, Swift 6, and any AirPods that expose head motion. Licensed under Apache 2.0.

Feynman is a CLI‑first, open‑source AI research agent that can search papers, web, and many scientific databases, then synthesize briefs, literature reviews, paper rankings, audits, replication plans, and ML‑training recipes. It ships as a self‑contained native bundle (or npm package) and optionally runs a local web workbench for notebooks, artifact previews, and compute orchestration. Built on the Pi agent runtime and AlphaXiv paper engine, it supports both hosted LLMs and local models via LM Studio, Ollama, LiteLLM, etc.

VoiceprintRecognition‑Pytorch is a PyTorch‑based speaker‑verification toolkit offering many modern backbones, pooling layers, loss functions and front‑ends, with scripts for data prep, training, evaluation and ready‑to‑use web/WeChat demos.

MiniMax Code is a terminal‑based AI coding assistant. It lets you chat with a LLM (MiniMax or any OpenAI/Anthropic‑compatible provider) to read, modify, and test code directly from the command line. Features include an interactive TUI, headless CLI for scripts/CI, custom model support, built‑in search and tool execution, session management, and editor integration via ACP. Install via a one‑click script or npm, log in (or use your own API key), and start issuing prompts like “fix the failing test”. The repo contains the source for the CLI (MIT‑licensed) and full documentation.

5ire is a free, cross‑provider AI chat assistant that uses the open Model‑Context‑Protocol (MCP) to attach tools (file system, databases, APIs, etc.) and includes a local multilingual vector store for RAG, a prompt library, bookmarks, and usage analytics.

An E-ink bird frame for Raspberry Pi that uses local AI to detect birds by audio and renders them as hand-cut 1800s natural history illustrations.

An audio-to-MIDI transcription plugin for DAWs that uses the MuScriptor transformer model to convert audio recordings into editable MIDI notes locally.

A GUI and API wrapper for OpenAI Whisper that provides asynchronous transcription services with a built-in editor for correcting text.

MT3 is a multi-instrument automatic music transcription model that converts audio recordings into musical notation using the T5X framework.

self‑learning‑skills is a meta‑skill that lets AI coding agents capture and reuse the procedural knowledge they discover during a session (e.g., deployment steps, secret‑lookup locations). It works with Claude Code, Cursor, and any tool that reads an `AGENTS.md` file, persisting “golden paths” as new skill files or notes and auto‑loading them in future sessions. Installation is a one‑liner via `npx skills`, a Claude Code plugin, or manual copy. The project follows the open Agent‑Skills standard and is MIT‑licensed.

ESP‑Skainet is Espressif’s on‑device voice‑assistant SDK for ESP32 chips. It bundles a tiny wake‑word engine (WakeNet), an offline command recognizer (MultiNet) supporting up to 200 languages, and a full audio front‑end (AEC, VAD, NS). The stack runs on ESP32‑S3 (or other ESP32 boards) with minimal RAM/flash, enabling privacy‑preserving voice control for smart‑home, wearables, or robotics. The repo includes ready‑to‑flash examples, detailed docs, and support for custom wake words and commands.

VocoType‑CLI is an open‑source, offline speech‑to‑text command‑line tool built on FunASR. It runs on CPU, supports Chinese‑English mixed transcription, offers a custom dictionary and AI‑based post‑processing, and can optionally use Volcengine’s cloud recognizer. The repo provides clear Python‑based installation steps, dataset export, and links to a full GUI version for non‑technical users.

Tel‑Agent is an open‑source gateway that connects SIP phone lines (and 24+ messaging platforms) to an AI agent you host yourself. It streams speech‑to‑text, LLM generation, and text‑to‑speech in real time, can call external HTTP services, records and transcribes every interaction, and runs on Windows/macOS/Linux or Docker with AGPL‑3.0 licensing.

CAD‑MCP is a Windows‑only Python server that exposes AutoCAD/GstarCAD/ZWCAD functionality through the Model‑Context‑Protocol. It lets LLM‑powered clients control drawing, querying, editing, layer, block, and file operations via 35 structured JSON tools, enabling natural‑language‑driven CAD automation.

Ackem is a Windows‑only Electron app that turns any OpenAI‑compatible LLM into a persistent, local‑first AI companion. It offers chat, searchable memory, emotional/relationship tracking, optional desktop pet, and an extensible plugin system, while keeping all data (including API keys) on the user’s machine. The project is actively maintained, released under AGPL‑3.0, and requires no Node.js runtime for end‑users.

whip is a Go‑based command‑line harness that runs an LLM‑driven coding agent. It streams model responses, calls tools (bash, file edits, sub‑agents) in parallel, and works with any OpenAI‑compatible endpoint. The tool auto‑discovers models from providers like OpenRouter, supports login‑based Codex subscriptions, offers built‑in terminal themes, and can be installed with a single script or via `go install`.

CatchMe is a local, privacy‑first tool that continuously records your computer activity (mouse, keyboard, windows, clipboard, screenshots, etc.), organizes it into a hierarchical activity tree, adds LLM‑generated summaries, and lets you query the whole history via CLI, a web dashboard, or an AI‑agent skill—without needing vector embeddings.

XLB is a Python library for differentiable Lattice Boltzmann fluid simulations. It runs on CPUs, single‑GPU (Warp), multi‑GPU (Neon), and TPU (JAX) hardware, offers grid‑refinement, many LBM models and boundary conditions, and integrates with JAX‑based ML tools for gradient‑based optimisation and physics‑informed learning.

A Node‑based CLI that lets you capture a DeepSeek Harness (runtime, ordered plugins, config patch, env‑var declarations) as an immutable, versioned *Preset Release* on the public DSH Plugin Hub. Others can fetch and apply the preset with a single command, getting exact plugin versions, load order, and runtime pinned. The tool supports planning, atomic staged installs, diff/upgrade/rollback, and integrates with AI agents via a safe plan‑apply workflow. All releases are content‑addressed (sha256), secrets stay local, and the CLI is MIT‑licensed.

Creator Agent is an open‑source, privacy‑first web prototype that lets a creator upload text, markdown, or video + subtitle files, approve them, and expose a deterministic, citation‑rich chat agent. It uses Auth0 for authentication, PostgreSQL for durable workspace metadata, and a zero‑AI local retrieval engine, with a documented “bring‑your‑own‑agent” HTTP contract for future model integration.

DevCodex is a Node‑JS CLI that orchestrates AI‑coding assistants (Copilot, Claude Code, etc.) with an intent‑driven, project‑aware workflow. It builds a structured project Profile, loads domain‑specific Skills on demand, persists task state across sessions, and requires evidence before marking work complete. Designed for developers and teams who need reproducible, auditable AI‑generated code changes.

Mira is an open‑source, self‑hosted AI code‑review service. It works with any LLM (OpenRouter, Ollama, AWS Bedrock, Codex CLI, etc.), stores diffs, embeddings, CVE data and usage logs in your own SQLite/Postgres, and presents results through a private dashboard. Features include noise‑filtered comments, a learning loop that turns rejected feedback into custom rules, full‑repo indexing, vulnerability alerts, and support for GitHub, GitLab and Forgejo. Deploy via Docker (one‑click Railway, Fly.io, Render) and configure per‑repo with `mira.yaml`/`.mira.yaml`. Licensed under Apache 2.0.

An AI‑agent skill that lets Claude Code or Codex understand natural‑language Unity tasks and automatically generate the corresponding Unity project code and assets. Install via marketplace commands or manually, works with Unity 6+, and is licensed under Unity’s Companion License.

A skill‑bundle for AI coding agents that adds classic lateral‑thinking methods (random stimulus, SCAMPER, Six Hats, etc.) with built‑in “abandonment” rules, letting the agent diagnose why ideas feel stale and generate structured, non‑obvious product directions.

QGIS MCP is an open‑source bridge that lets AI agents (Claude, Gemini, Codex, etc.) control QGIS via the Model Context Protocol. It provides a plugin inside QGIS and an external FastMCP server exposing 118 GIS‑focused tools (layer handling, editing, processing, rendering, layouts, etc.). Install the plugin from QGIS, add a tiny `uvx`‑based server to your LLM client, and you can create projects, edit features, run processing algorithms, and render maps purely through AI‑driven commands.

kiwi‑mem is a self‑hosted FastAPI gateway that adds human‑like long‑term memory to LLM chats. It stores facts in PostgreSQL + pgvector, assigns each a heat that decays over time, merges related pieces during a nightly “Dream” cycle, and compresses old chats into hierarchical calendar summaries. The system works with any OpenAI‑compatible or Anthropic native API, provides a web admin panel, and runs via a single Docker‑Compose command.

OpenLive is an open‑source desktop app that adds on‑device speech‑to‑text, text‑to‑speech (with optional voice cloning), and optional camera/screen capture to any LLM or coding‑agent you choose. It runs the entire audio pipeline locally (WebGPU), sends only the transcript to the model, and returns spoken responses, enabling private, fee‑free voice‑driven AI assistants.

A music recommendation system that combines acoustic feature encoding via MLP and collaborative filtering via Track2Vec to provide high-quality, genre-aware song suggestions.