nvidb
See the GPUs on all your SSH servers in one terminal view, and queue jobs onto whichever has room.
pip install nvidbOpen source, shipped tools only
These are the open-source tools I have built and released: for GPU servers, remote development and coding agents, desktop configs, LaTeX, odd hardware and campus life. Each one comes with a way to install or open it. Research prototypes stay off this page.
25 projects on 7 shelves, 58 stars on GitHub
Tools that came out of running shared GPU servers: see what is busy, split a card fairly, and develop CUDA code on a machine with no GPU at all.
See the GPUs on all your SSH servers in one terminal view, and queue jobs onto whichever has room.
pip install nvidbRun CUDA-facing code and plan GPU memory without a real GPU: fake devices for PyTorch, NVML and NCCL calls, plus memory estimators for LLM training and serving.
Linux wheel on Releases →Split one NVIDIA GPU's SMs and memory between jobs, give one of them priority, and rebalance while they run.
pip install nvertakeAn nvtop-style terminal monitor for the GPU and Neural Engine on Apple Silicon Macs.
pip install metopFor working on remote machines and alongside coding agents: providers, skills, shells and files, without setting up the server first.
A desktop panel for switching Claude Code providers on remote SSH servers, compatible with cc-switch.
Download for macOS, Windows or Linux →Keep the skills of Claude Code, Codex, Gemini and other coding agents in one store, and switch whole skill sets at once.
go install github.com/FanBB2333/skim/cmd/skim@latestTurn a shell you already have (tmux, a web terminal, docker exec, ssh -T) into file and command tools for an agent, with an MCP server.
go install github.com/FanBB2333/ptyrelay/cmd/ptyrelay@latestMount a remote code-server workspace as a local folder over NFS or WebDAV, with only its URL and password: no SSH needed.
go install github.com/FanBB2333/code-local/cmd/code-local@latestRun Linux VMs on a Mac with different CPU, memory, network and browser setups, and audit the device details each guest can see.
Desktop apps that replace hand-edited config files and clunky web consoles.
Edit Surge profiles with structured editors, a live preview of the .conf and a tester that shows which rule a URL matches.
See it on the showcase site →A desktop browser for Alibaba Cloud OSS: tabs, an editable oss:// address bar, previews and an upload/download queue.
Download for macOS or Windows →Reading other people's LaTeX, and writing your own outside the browser tab.
A Chrome extension that opens any arXiv paper's TeX source in a VS Code-style viewer, with a file tree and highlighting.
Add to Chrome →Mount Overleaf projects as a local folder and edit them in any editor, or with Claude Code; changes sync back on save.
go install github.com/FanBB2333/downleaf/cmd/downleaf@latestGetting unusual hardware to talk to a Mac.
Use the first-generation DJI 4G dongle as a USB modem on macOS, with an AT console, SMS and eSIM tools.
Build from source →A macOS app for reading and writing Mifare Classic cards with a PN532 NFC module.
Build from source →Small fixes for university systems: subject registration, VPNs that drop, and waiting for results.
A Chrome extension that makes PolyU subject registration searchable, with timetables, vacancies and clash checks inline.
Build and load in Chrome →The PolyU GlobalProtect VPN in a Docker container, with a web panel for the MFA login and a local SOCKS5 proxy.
git clone https://github.com/FanBB2333/PolyGP.git && cd PolyGP && cp .env.example .env && docker compose up -d --buildA desktop app that watches ZJU thesis blind-review results and sends a Bark notification when they change.
Download for macOS or Windows →A watchdog that signs a dropped aTrust VPN session back in, by looking at the remote desktop over SSH and VNC.
pip install -r requirements.txt && python3 auto_atrust.pySkills that change how a coding agent reasons or writes. They work with Claude Code, Codex and other agents that read SKILL.md files.
Fourteen schools of philosophy, from Daoism to Popper, turned into planning and checking strategies, with a guide for picking one per task.
The model grills its own draft for gaps and jargon first, then explains the answer from first principles.
Debug in four diagnostic phases, borrowed from traditional Chinese medicine, before writing a minimal fix with a test.
Strips corporate jargon, filler openers and flattery from GPT models' Chinese output.
The opposite direction: gives Claude, Gemini and other models GPT's chatty Chinese style.
A joke skill: a cat walks across the keyboard and leaves harmless mischief in the agent's code.
To use one, copy its folder into your agent's skills directory, for example .claude/skills/<name>/ for Claude Code.