Plug in a USB. Offline AI coding with Opencode + Ollama. Works on Windows, macOS & Linux.
Plug in a USB. Get an offline AI coding agent on any laptop. One stick. Six targets — Windows (x64 + ARM64), macOS (Apple Silicon + Intel), Linux (x64 + ARM64). Zero install on the host.
@user_50c8d6f4npx code-stick install
You want an AI coding agent that works on:
Cloud agents won't work. Installing Ollama + a 5 GB model on every machine you touch won't either. code-stick is the in-between: install once on a USB, then run the agent from the stick on whatever laptop is in front of you. Host stays clean.
Under the hood it's opencode (terminal coding agent) talking to Ollama (local model runner), both pre-built for every target, both serving from the USB on 127.0.0.1:11434. Quit the agent, the Ollama process dies, the host has no residue.
npx code-stick install # auto-detects USB; pick a model interactively
Plug the stick into any supported machine, double-click the launcher for that OS:
start-windows.batstart-mac.commandstart-linux.shopencode runs in the terminal talking to a USB-local Ollama on
127.0.0.1:11434. Quitting opencode kills the Ollama process. Nothing is
left behind on the host.
The recommended default is Qwen2.5-Coder 7B — fits a 32 GB stick, runs on 8 GB RAM, best all-rounder for code. Other curated options scale up to 32B for serious refactors on larger sticks.
| Model | Size | Stick / RAM | Best for |
|---|---|---|---|
| Qwen2.5-Coder 7B ⭐ | ~4.7 GB | 32 GB / 8 GB | All-rounder for coding |
| Qwen2.5-Coder 14B | ~9.0 GB | 64 GB / 16 GB | Stronger reasoning + multi-file |
| Qwen2.5-Coder 32B | ~20 GB | 128 GB / 32 GB | Top-tier OSS coder, near-frontier |
| Phi-3 Mini 3.8B | ~2.2 GB | 32 GB / 4 GB | Lightweight, low-spec hardware |
Full curated list, BYO Ollama tag instructions, and per-model context
windows: docs/MODELS.md.
⚠ USB 2.0 sticks are unusable for models >7B. ~30 MB/s read means a 20 GB model takes ~11 minutes just to warm up. Use USB 3.0 or better for anything bigger than 7B.
⚠ First-prompt latency on large models can be 30–90 s — Ollama is mmap'ing the weights off the USB. Subsequent prompts are fast once the OS page-caches the blob.
⚠ macOS first-launch: Gatekeeper will block the unsigned launcher (notarization is on the roadmap). Right-click
start-mac.command→ Open → Open. Details:docs/TROUBLESHOOTING.md.
| Command | Description |
|---|---|
code-stick install | Set up code-stick on a USB |
code-stick start | Start opencode + Ollama from a USB |
code-stick status | Show what's installed |
code-stick doctor | Live audit (port + Ollama + opencode + model store) |
code-stick prune | Run ollama prune on the USB model store to reclaim orphaned blobs |
code-stick update | Refresh launcher scripts and manifest timestamp |
code-stick upgrade-engine | Re-download Ollama + opencode without nuking the model store |
code-stick add-model [id] | Pull another model onto the stick |
code-stick remove-model [id] | Remove a model from the stick |
code-stick add-targets [list] | Add OS targets to a stick installed with --targets (restore portability) |
code-stick uninstall | Wipe code-stick from the stick |
Full flag reference, install modes (Fast vs Direct), and --targets
trimming: docs/COMMANDS.md.
--targets host (see
docs/STORAGE.md); 16+ GB for full portability with
small models; 32+ GB (recommended default + Qwen 7B); 64+ GB (medium), 128+
GB (large).docs/MODELS.md.v0.2.1 — opencode v1.x, per-model num_ctx bake. Full install flow on
Windows (x64 + ARM64), macOS, Linux (x64 + ARM64) — tested manually per
target. Surface Pro / Snapdragon X laptops boot natively (ARM64) with
x64-via-Prism fallback if only the x64 target is staged. Docker-based
Linux smoke test in CI. No telemetry, no background HTTP. Bug reports
are local-only and redacted before write.
Known gaps: macOS not yet notarized, no Windows/macOS CI runners.
File an issue or open a PR if you hit anything: github.com/MuhammadUsmanGM/code-stick/issues.
docs/STORAGE.md — USB sizing, default 6-target download (~5–7 GB binaries), 8 GB + --targets hostdocs/TROUBLESHOOTING.md — hallucinations, Gatekeeper, MAX_PATH, port conflicts, bug reportsdocs/MODELS.md — full model catalog, BYO Ollama tags, quantization, context windowsdocs/COMMANDS.md — every subcommand and flag, Fast vs Direct, --targetsdocs/ARCHITECTURE.md — stick layout, runtime behavior, process isolationdocs/SECURITY.md — trust model, threat boundariesdocs/TRUST.md — supply-chain provenancenpm install
npm run typecheck # tsc --noEmit
npm test # vitest run (unit + launcher snapshot tests)
npm run build # tsup → dist/cli.js
npm run smoke:docker # full end-to-end install + launch in Linux Docker
MIT — Muhammad Usman (github.com/MuhammadUsmanGM)
CLI proxy that reduces LLM token consumption by 60-90% on common dev commands. Single Rust binary, zero dependencies
@user_b7c5413dTUI Interface to Local and cloud LLM providers
🎨 Local-first, open-source Claude Design alternative. 🖥️ Native desktop app. ⚡ 259+ Skills · ✨ 142+ Design Systems 🖼️