Open Source AI

Open-source, local-first AI voice studio — clone a voice from a few seconds of audio, generate speech across seven TTS engines, dictate into any app, and give MCP-aware agents (Claude Code, Cursor, Cline) a voice of their own. A free, on-device alternative to ElevenLabs and WisprFlow for macOS, Windows and Linux. MIT-licensed.

An open-source JavaScript in-page GUI agent from Alibaba. Drop a single script tag on any website and it adds an AI assistant that reads and controls the page through the DOM from natural-language instructions — clicking, filling forms, navigating and explaining as it goes. Works with any OpenAI-compatible model (Qwen/DashScope, OpenAI, DeepSeek, Gemini, Ollama) and ships a free demo model so you can try it instantly. See the attached user guide for a step-by-step, toggle-launcher implementation.

A self-hosted personal AI assistant that lives in the apps you already use — WhatsApp, Telegram, Slack, Discord, iMessage and more — running locally on your own machine around the clock. Message it from your phone to manage calendar, email, research and tools with your own model key. `npm install -g openclaw@latest`.

An agent-harness optimization system for Claude Code (also Codex, OpenCode, Cursor): adds skills, persistent memory, security scanning and a research-first, test-before-you-ship workflow, so the agent stops committing broken code and remembers your project between sessions. Handy commands include /ecc:plan, /analyze-repo and `npx ecc-agentshield scan`.

"Skills for real engineers, not vibe coding" — the skills Matt Pocock uses daily, straight from his .claude directory. They make Claude less reckless: ask sharp questions before building, do proper test-first development, and keep code maintainable. Leans TypeScript/JavaScript. Install with `npx skills@latest add mattpocock/skills`.

Garry Tan's opinionated Claude Code setup — 23 tools that turn one agent into a virtual startup team: CEO, engineering manager, designer, reviewer, QA and release manager. Run work through a sprint with commands like /office-hours, /plan-eng-review, /review, /qa and /ship.

Turns any folder — code, SQL schemas, shell/R scripts, docs, papers, even images or video — into a queryable knowledge graph, so your coding agent queries a map instead of re-reading every file each session. Cuts token usage sharply on large codebases. Works with Claude Code, Codex, Cursor, Gemini CLI and more.

ByteDance's open-source long-horizon "SuperAgent" harness for big, multi-step jobs (deep-research reports, dashboards, slide decks). A lead agent splits the task across parallel sub-agents that run real code in a sandbox with their own memory and filesystem, then stitches the deliverable together. Docker-based and model-agnostic.

Vercel Labs' browser-automation CLI for AI agents — lets a coding agent drive a real browser (navigate, click, fill forms, extract data) straight from the command line, so agents can use the web the way a person would.

Open-source agentic video production for your AI coding assistant: describe a video and it researches, scripts, narrates, sources or generates visuals, edits the timeline and renders a finished MP4. 12 pipelines (cinematic trailers, explainers, real-footage documentaries) with a free local path. Needs Python, Node.js and FFmpeg.

Garry Tan's opinionated agent "brain" — a memory layer for people, companies, meetings and notes that synthesizes cited answers (e.g. "what should I know before my meeting with Alice?") instead of just returning links. Runs locally in about 30 seconds, or scales to a full auto-updating stack. Plugs into Claude Code via MCP.

817 structured cybersecurity skills for AI agents across 29 domains — malware analysis, threat hunting, cloud breaches, incident response — each a playbook mapped to frameworks like MITRE ATT&CK, NIST CSF and MITRE ATLAS. Install only the domain you need to avoid context bloat. `npx skills add mukul975/Anthropic-Cybersecurity-Skills`.

From the browser-use team: edit videos with coding agents. Point your AI agent at footage and drive edits programmatically, bringing the browser-use agent approach to video editing.

NVIDIA's security scanner for AI-agent skills. Point it at a skill before installing and it flags hidden instructions, prompt injection, data exfiltration and over-broad permissions, returning a 0-100 safety score. A cited study found roughly a quarter of skills had a vulnerability — scan first, install second.

The open-source managed agents platform — turn coding agents into real teammates: assign tasks, track progress, and compound skills.

Hugging Face's open-source framework for building local voice agents. It chains four swappable stages — voice detection, speech-to-text, an LLM, and text-to-speech — into a low-latency conversation loop, and exposes an OpenAI Realtime-compatible WebSocket API so existing Realtime clients work unmodified. Every stage is picked by flag (Whisper or Parakeet, a hosted or llama.cpp model, Qwen3-TTS or Kokoro), and the whole thing runs offline on CPU, CUDA, or Apple Silicon — no per-minute cost, no audio leaving your machine. Production-tested on thousands of Reachy Mini robots. Apache 2.0.

Hold up both hands, frame a box with your fingers, and a live AI-generated world appears inside it — regenerated at 30fps by Decart Lucy 2.5 (realtime video-to-video over WebRTC), so it moves with you: blink and it blinks, wave and it waves. Hand tracking is MediaPipe Hand Landmarker, compositing is plain Canvas 2D, and the whole thing is a single static page you serve with a one-line Python http.server. Keys 1–6 swap live style prompts (3D Movie, Anime, Cyberpunk, Watercolor, LEGO, or your own) with no reconnect. Bring your own Decart key; without one it falls back to a local colour filter so the tracking is still demoable.

Real-time face swap and one-click video deepfake from a single source image. Point it at a webcam, video, or photo and it swaps faces live — with a mouth mask that keeps the original mouth movement, face mapping to handle several people at once, and optional GFPGAN face enhancement. Built on InsightFace and the original ROOP project, running ONNX models on CPU or accelerated via NVIDIA CUDA, AMD DirectML, Intel OpenVINO, or Apple CoreML. AGPL-3.0. The project asks users to get explicit consent before using a real person's face and to clearly label output as a deepfake; it ships built-in checks that block inappropriate media. Note the InsightFace models it relies on are licensed for non-commercial research use only.