About this project

Step Code is a terminal-based AI coding agent developed by StepFun AI. It handles the full development task loop—reading code, making changes, and running tests—while operating autonomously through long-running task delegation. Key capabilities include token-efficient execution tuned alongside Step models, with long-horizon tasks split across parallel subagents that maintain isolated contexts to avoid redundant content in the main conversation. It supports MCP servers, Agent Skills, and plugins out of the box, with most Claude Code plugins directly compatible. A built-in StepPage plugin allows publishing local directories as shareable static websites with version management and rollback via a single command. The agent provides interactive TUI and headless modes. Tasks can be delegated to `/goal` for autonomous execution or scheduled with `/cron`. Users can switch between Step models via `/model` after signing in with a Step Plan account or API key. Multiple permission modes (Ask, Read Only, Bypass, Autopilot) control command execution safety. Installation is available via platform-specific installers for macOS, Linux, WSL, and Windows (beta). The tool can also be built from source using Git, Node.js, and pnpm. Sessions can be resumed later, and an `/init` command generates project guides compatible with AGENTS.md format.