AI Deck

Command Code — A Terminal-Resident AI Coding Agent That Learns Your Style from Every Accept, Reject, and Edit

An AI coding agent that lives in your terminal. Its defining feature is the proprietary “taste-1” model, which treats every accept, reject, and edit of a suggestion as a learning signal, gradually picking up the conventions specific to each project and each developer. The service grew out of a pivot by Langbase, a serverless AI platform, and was founded by Ahmad Awais. It has raised $5M in a seed round led by PWV, the fund of GitHub co-founder Tom Preston-Werner. By design, your code stays local, and the “taste” it learns is kept on your machine as well.

Key Features

  • Continuous learning with taste-1: Whether you accepted a suggestion, rejected it, or reworked it is treated as a signal, with the goal of sparing you from repeating the same correction. The accumulated conventions are reused as project-level skills
  • Three execution modes: An interactive CLI, a headless mode callable from scripts (-p / --yolo), and a background sandbox that runs without interrupting your current work
  • Context that persists across sessions: Define agents with /agents and accumulate your personal assumptions and decisions in /memory so they carry over into the next session
  • Extensible via skills, commands, and MCP: Package routine work into /skills and /commands, and connect to external services or your local environment through /mcp. A plugin mechanism is also available
  • Support for 50+ models: Alongside Anthropic, OpenAI, and Google, you can choose from a wide range including DeepSeek, Qwen, MiniMax, and GLM. BYOK, using your own API key, is supported as well
  • Team sharing: Share sessions with /share, or hand learned conventions between teammates with npx taste push / pull

To get started, install globally with npm i -g command-code@latest, authenticate in the browser via cmd login, then run cmd inside a project directory. Node.js 22 or newer is required. An /import command is provided for bringing over existing configurations from Claude Code, Codex, Cursor, Gemini, and others.

Pricing

Pricing is credit-based: each plan differs in the monthly credit allowance granted and the range of models available.

PlanMonthly PriceCredits GrantedMain Features
Go$1$10 worthtaste-1 and select premium models, Discord support
GOAT$10$70 worthPer-model allowances across 29+ models, usage analytics
Pro$20$80 worthOpen-source and premium models, access to higher-end models
Max 10×$100$150 worthHigher rate limits, priority support
Max 20×$200$300 worthAll models, the highest rate limits
Provider (API)$15 + pay-as-you-goOpenAI/Anthropic-compatible endpoints, per-request analytics
Teams$40PooledShared credits and consolidated billing
EnterpriseContact for pricingCustomUnlimited seats, dedicated support, SLA

Unused credits are said to roll over, and you can cancel at any time. No discounted annual plan is listed on the official pricing page. Time-limited promotions targeting specific models appear from time to time, which can make the amount you can effectively spend larger than the granted credits.

Pricing is as of August 2026. Please check the official site for the latest pricing.

Pros & Cons

Pros

  • Because it learns your preferred style, the effort of repeating the same correction steadily decreases
  • Interactive, headless, and background modes let you cover everything from hands-on work to automation with one tool
  • With 50+ models to choose from and BYOK support for your own API key, you are not locked into a narrow set of models
  • Your code is not used for training, and the conventions it accumulates stay local
  • The cheapest plan starts at $1 per month, making it easy to try

⚠️ Cons

  • No fully free plan appears on the pricing page, so continued use assumes a monthly subscription
  • The credit system means you may hit your allowance within the month depending on usage, making actual costs harder to predict
  • It is built around the terminal, which may not suit people who work primarily in a GUI editor
  • As a service that launched in 2026, there is still less information and fewer case studies than for Claude Code or Cursor
  • The benefits of learning emerge the more you use it, so a short trial may not reveal the difference

Comparison with Similar Services

CriteriaCommand CodeClaude CodeCursorOpenCode
ProviderCommand Code (formerly Langbase)AnthropicAnysphereOpen-source community
Form factorTerminal CLITerminal CLIEditor (VS Code-based)Terminal CLI
Learning your styleContinuous learning from accept/reject/edit via taste-1Memory and config filesRule filesConfig files
Models available50+ (multi-vendor, BYOK supported)Claude-centricMultiple vendorsMultiple vendors (BYOK)
Pricing approachCredit-based (from $1/month)SubscriptionSubscriptionFree core + pay-per-model

Who Is It For

  • Developers who keep making the same corrections to AI suggestions and want to reduce that fixing work itself
  • People who work mainly in the terminal and want to delegate to an agent without leaving it
  • Anyone who does not want to be locked to a single model vendor, or who wants to reuse API keys they already hold
  • Development teams looking to align coding conventions and stylistic habits across members
  • People who want to invoke an AI agent headlessly from CI or scripts

Summary

Command Code is a coding agent that goes after the overhead of explaining your preferences to an agent over and over. The continuous learning of taste-1 and its local-retention design pay off most for developers who stay with the same project for a long time. On the other hand, there is no fully free tier, and the credit model takes some experimentation before you get a feel for the cost. A reasonable approach is to start with the $1/month Go plan to see whether it fits your workflow, then consider a higher tier or the Provider plan once your usage becomes predictable.

← Blog