Command Code

Command Code is a coding agent for working in a software project, with CLI and desktop entry points. Its documentation describes interactive, headless and planning modes, model choices, skills and MCP connections. The product also presents a preference-learning system called Taste for adapting suggestions to how a developer writes code.

Start inside a project

The documented CLI setup installs the command-code package, then starts cmd in a project directory. The docs list workflows for planning, refactoring, pull-request review, commits, tests and documentation. These are task entry points, not evidence that a particular change will pass review or preserve behavior.

Begin with a bounded issue and inspect the agent's file edits and command output. Keep the repository's existing test and review process. A generated regression test needs the same scrutiny as an application change, including whether it would have caught the original defect.

The homepage depicts a startup-paint investigation with shell commands, file reads, browser checks and background tasks. Its displayed test counts and timing are a product example, not a measured result for another repository.

Choose a model and execution setup

Command Code lists open-model families alongside other providers. Its docs also advertise temporary model offers and boosted allowances. Confirm current availability, billing and the chosen model's capabilities before making an offer part of a project's operating budget.

Little-Coder is another terminal-agent option for smaller local models. Its documented local-server setup is a useful comparison if local inference is the deciding requirement; Command Code's open-model support alone does not prove that a session runs offline.

MCP connects tools and services, and skills provide reusable commands. Inspect those connections and their permissions before using them in an unfamiliar project. The retrieved overview does not establish a complete sandbox or approval policy for every mode.

Review what the agent learns

The Taste description says accepted, rejected and edited suggestions shape coding preferences. It also describes sharing those preferences with a team. That may help maintain project style, but style agreement does not establish correctness or security.

Check what preference information is stored, how it can be inspected or removed, and what sharing exposes before applying it across client repositories. The captured pages do not provide complete retention or privacy details. Use the documented workflows to assist development while keeping code review, permission boundaries and release decisions under the team's control.