Command Code

Command Code is a frontier AI coding agent that lives in your terminal and continuously learns your coding taste through a proprietary system called taste-1. Founded by Ahmad Awais and backed by $5M in funding, the tool installs via npm (npm i -g command-code) and works across any project without requiring IDE changes. The homepage states the company has shipped 252 releases, with the current version at v0.30.0.

What It Is

Command Code is a CLI-first coding agent in the same category as Claude Code and GitHub Copilot, but differentiated by its continuous reinforcement learning system. Every accept, reject, and edit a developer makes becomes a signal that the agent distills into project-level /skills and personal /memory. The result is an agent that adapts to individual coding conventions — naming patterns, toolchain preferences, test frameworks — rather than generating generic output. The underlying system is described by the company as "taste-1," a meta neuro-symbolic AI model that pairs LLMs with learned developer preferences.

How the Taste System Works

The taste-1 system operates through three mechanisms:

• Continuous learning: Every interaction (accept, reject, edit) is recorded and auto-generates project-level skills stored in .commandcode/taste/
• Portable taste packages: Developers can run npx taste push to publish a skill and npx taste pull to install one, enabling team-wide taste sharing
• Onboarding from other agents: The tool can import existing preferences from Claude Code, Cursor, or other agents with a single command, and can also distill taste from any local path or GitHub repo

Taste data is stored locally on the developer's machine. The company states it never trains on user code.

Modes and Agentic Tools

Command Code supports multiple operating modes to fit different workflows:

• Interactive mode: Real-time conversational interface for chat-driven development
• Headless mode (-p, --yolo): Non-interactive single-query mode for CI/CD pipelines and scripts
• Plan mode: Read-only analysis without executing changes
• Auto-accept mode: Skip confirmation prompts, toggled with shift+tab
• Background sandbox: Fire-and-forget isolated tasks that run to completion and report back

Built-in tools include file operations, shell execution, extended thinking, agentic /review for PRs, web search and fetch, todo tracking, MCP server integration, custom slash commands, sub-agents, and checkpoints with rewind capability.

Model Access and Provider Flexibility

The agent supports a broad range of frontier and open-source models out of the box, including Anthropic (Claude Opus 4.8, Sonnet 4.6, Haiku 4.5), OpenAI (GPT-5.5, GPT-5.4, Codex), DeepSeek (V4 Pro, V4 Flash), Kimi (K2.6, K2.5), GLM (GLM-5.1, GLM-5), Qwen, and MiniMax. BYOK (bring your own key) is supported for any provider. The company positions open-source model support as a core differentiator, with the harness designed to enable reliable tool-calling on open models.

Update: v0.30.0 and 252 Releases

The changelog shows v0.30.0 as the current release, with 252 total releases shipped. Recent changes listed on the homepage include: repair of read_file/edit_file/glob aliases, a two-step /model picker for model and reasoning effort selection, a new /effort command to set reasoning effort per model, and a fix preventing "Thinking…" indicators from getting stuck. The pace of releases signals active, daily development consistent with the company's stated "Ship. Daily." value.

Team and Backing

The About page lists the founding team led by Ahmad Awais (Founder & CEO), with backgrounds spanning Google, NASA, GitHub, Microsoft, Cloudflare, and the Linux Foundation. Investors listed on the About page include Preston-Werner Ventures, Firestreak Ventures, Mento VC, Alumni Ventures, and AltaIR Capital, alongside individual angels including Tom Preston-Werner (GitHub co-founder), Amjad Masad (Replit CEO), Luca Maestri (Apple CFO), and Guy Podjarny (Snyk/Tessl CEO). The company is headquartered at 2261 Market St, San Francisco, CA.