3 November 2026

Computer use CLI for AI agents via the OS accessibility tree

A CLI and C library that reads and operates desktop apps through the OS accessibility tree, with no pixel matching.

• A snapshot returns JSON with a ref for every element an agent can act on.
• Available for macOS, Windows, and Linux; macOS needs Accessibility permission.

Cross-session memory for many coding agents

Captures what an agent does during sessions, compresses it with AI, and injects relevant context into later sessions.

• Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, and OpenCode.
• Captures tool usage observations and generates semantic summaries from them.

Web scraping framework from one request to a full crawl

A Python library that combines fast HTTP requests, stealth browsers, parsing, and concurrent crawling.

• Fetchers support JavaScript rendering and Cloudflare Turnstile handling.
• Concurrent spiders offer proxy rotation, retries, pause/resume, and JSON, JSONL, CSV, XML exports.
• Usable from Python, the terminal, or an AI agent through its MCP server.

Independent, customizable tiling Wayland compositor

A dynamic Wayland compositor with animations, blur, and many customization options, without relying on wlroots.

• 100% independent: no wlroots, libweston, kwin, or mutter.
• Has plugin support, a built-in plugin manager, and socket-based IPC.
• Config reloads instantly when the file is saved.

Turn Claude Code into a game studio with 49 agents

Gives a single Claude Code session the structure of a game studio: specialized agents with defined roles, escalation paths, and quality gates.

• 49 agents organized in three tiers: directors, department leads, and specialists.
• Ships with 15 validation hooks, 14 coding rules, and 62 document templates.

Run large MoE models on consumer hardware

A Mixture-of-Experts serving engine for large open-weight models across the GPU, CPU, and memory of a personal PC.

• Co-executes on CPU and GPU with a global LRU expert cache.
• Offers Anthropic- and OpenAI-compatible APIs for agents such as Codex and Claude Code.
• Supports NVIDIA RTX 30, 40, and 50 series GPUs.

Control surface for the coding agents on your machine

Controls the agents installed on your computer through mobile, web, and Electron desktop apps.

• Uses your existing subscriptions on Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity, and fx.
• Available on iOS, Android, web, and desktop; the t3 command starts a local server.

Foundation model for the language of financial markets

A family of decoder-only models pre-trained on candlestick (K-line) sequences from over 45 global exchanges.

• A tokenizer turns OHLCV data into hierarchical discrete tokens, then an autoregressive Transformer is trained on them.
• The mini, small, and base models are available on the Hugging Face Hub.

Differentiable geometric computer vision library on PyTorch

A PyTorch-based computer vision library of differentiable image processing and geometric vision algorithms.

• Supports batch transformations, auto-differentiation, and GPU acceleration.
• States its direction as the reference implementation for differentiable vision in the PyTorch ecosystem.

Run huge MoE models in pure C, experts streamed from disk

A dependency-free C engine that keeps the active part of a model in RAM and reads the other experts from disk when needed.

• No GPU required, but it can use Vulkan or CUDA for GGUF models.
• Runs 13 engines, including GLM-5.2/5.3, DeepSeek V4 Flash, and Qwen3.6.

Small, composable agent skills for real engineers

A set of agent skills its author uses daily for real engineering, designed to be small, easy to adapt, and composable.

• Works with any model and does not take over your process.
• Installed as a Claude Code, Codex, or GitHub Copilot plugin, or copied in through skills.sh.

Git worktree CLI for parallel agent workflows

A CLI for managing Git worktrees so running several AI agents in parallel is as easy as using branches.

• Three core commands make worktrees as easy as branches.
• Has hooks to automate local workflows and copy-on-write build caches.

Open-source orchestration for teams of AI agents

A Node.js server and React UI that runs a team of AI agents: goals, tasks, costs, and permissions in one dashboard.

• Three steps: define the goal, hire the team, approve and run.
• Bring your own agents: Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and more.

Simple harness for training LLMs on a single GPU node

A minimal, hackable experimental harness covering every major LLM stage from tokenization to inference.

• A single --depth setting sets the model size; other hyperparameters are computed automatically.
• A GPT-2 capability model can be trained for about $48, roughly 2 hours on an 8xH100 node.

Training-free language model modification

A tool that changes language model behavior with mathematical techniques, without expensive training.

• Uses contrasting example prompts and matching scorers; the default action suppresses refusals.
• Other configs reduce slop, humor, and sycophancy.
• 4-bit bnb_4bit quantization reduces the VRAM needed.