Skip to content
MuhammadUsmanGM

MuhammadUsmanGM/code-stick

CLILive in production

> Plug in a USB. Get an offline AI coding agent on any laptop.

Plug in a USB. Get an offline AI coding agent on any laptop - Windows, macOS, Linux. opencode + Ollama, pre-built for 5 targets, zero install on the host.

57 1 since joining 12TypeScriptPush 22d agoListed 7d agoother

www.npmjs.com/package/code-stick

coding-agentlocal-llmoffline-aiollamaopencodeusb
  • TypeScript88.7%
  • JavaScript6.7%
  • Shell2.4%
  • EJS1.9%
  • Dockerfile0.3%
View on GitHub

Report a problem

1 Review

code-stick is a genuinely useful take on local AI tooling: instead of asking every machine to become a full Ollama/opencode workstation, it packages the runtime around the real constraint many developers hit, which is “I need this to work on whatever laptop is in front of me, possibly offline, without leaving the host dirty.” The README does a strong job explaining that use case in practical terms: airgapped environments, shared laptops, weak wifi, and privacy-sensitive code are all concrete scenarios where the USB-first design makes sense. I also like that the project is clear about its tradeoffs, especially model size, USB 2.0 performance, first-prompt latency, Gatekeeper friction on macOS, and the current lack of Windows/macOS CI.

The repo structure looks thoughtfully organized for a CLI: TypeScript dominates the codebase, src is split into catalog, commands, core, state, and utils, and the npm package exposes a clean code-stick binary with focused commands like install, start, doctor, add-model, add-targets, and upgrade-engine. The docs are a standout part of the project. Having separate pages for models, storage, commands, architecture, security, troubleshooting, and trust/provenance is exactly what this kind of tool needs, because users are being asked to run prebuilt binaries and large model artifacts from removable media.

The biggest improvement I would make is to turn the stated manual cross-platform testing into more visible automated confidence. The README already notes Docker-based Linux smoke testing and snapshot/unit tests with Vitest, which is good, but the project’s core promise is cross-OS portability. Even a small matrix that validates launcher generation and path handling on Windows, macOS, and Linux would make adoption easier. I’d also consider adding a short “threat model in one minute” section near the install command, linking to the deeper security/trust docs, since security-conscious users are the exact audience most likely to hesitate before running npx code-stick install.