What it does

Command Code is an npm-installed CLI tool (cmd) that generates code, fixes bugs, writes tests, and refactors projects. It wraps the usual AI-coding-agent workflow—prompt, generate, accept/reject—inside a terminal session with slash commands, bash mode (!), and file autocomplete (@).

The interesting bit

The hook is “Taste”: a persistent profile that supposedly learns from every accept, reject, and edit you make. The README calls this “taste-1” and frames it as “meta neuro-symbolic AI,” which is doing a lot of heavy lifting as a phrase. Portable via npx taste push/pull so teams can share accumulated preferences. Whether this actually compounds or just decays slower than prompt-engineering is unclear from the README alone.

Key highlights

• Interactive terminal UI with / commands, ! for shell escape, @ for file mentions
• Claims continuous learning from feedback signals (accept/reject/edit)
• Team taste sharing through npx taste push/pull
• Supports multiple frontier models (Anthropic, OpenAI, DeepSeek, GLM, Kimi per topics)
• Ships as a global npm package: npm i -g command-code

Caveats

• “Neuro-symbolic AI” and “taste-1” are branded terms with no technical detail in the README; actual mechanism is opaque
• Heavy reliance on external docs (commandcode.ai) for core concepts—README is essentially a landing page
• No benchmarks, comparison data, or open-source model weights shown

Verdict

Worth a spin if you’re already cycling through AI coding tools and want one that remembers your preferences. Skeptics should wait for someone to reverse-engineer whether “Taste” is more than fancy prompt prefixing.