SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Step Code - StepFun's Open-Source Terminal Programming Agent

Author: AI小集 · Original source: AI工具集

What is Step Code

Step Code is an open-source terminal launched by StepFunprogrammingagent, licensed under the MIT License, capable of directly completing code writing, debugging, execution, and delivery in the terminal. The agent tied for first place with an 80.9% pass rate on the Terminal Bench 2.1 evaluation, with the lowest token consumption among the tied entrants; it also ranked first on the self-built Multi-Frame long-horizon benchmark. Step Code has built-in StepPage one-click static site publishing capability, enabling development and delivery to be closed-loop in the terminal.

Step Code

Main Features of Step Code

  • Terminal Programming Agent: Directly completes the entire process of code writing, debugging, execution, and delivery in the terminal, with no need to switch development environments.
  • Efficient Long-Horizon Task Execution: By optimizing context management and the tool invocation chain, it reduces redundant information and ineffective calls, allowing long tasks to be completed stably.
  • Low Token Consumption: The lowest token cost among agents that meet the pass-rate target, significantly saving actual usage costs at the same capability level.
  • StepPage One-Click Publishing: After the local static page is built, one command can publish it as an accessible online website.
  • Multi-Platform Support: Supports macOS, Linux, and WSL, with a PowerShell beta version available for Windows.
Dig Into GitHub

Follow WeChat and reply “Open Source” to join theAI Open Source Project Communication Group

How to Use Step Code

  • Install: On macOS/Linux/WSL, run curl -fsSL https://static-openapi.stepfun.com/stepcode/install.sh | bash, and on Windows run with PowerShell irm https://static-openapi.stepfun.com/stepcode/install.ps1 | iex。
  • Verify Installation: After reopening the terminal, run step --version to confirm the installation was successful.
  • Log In: Run in the terminal step login, or enter after entering the TUI /login, log in to the Step Plan subscription through browser OAuth, or set STEP_API_KEY to use an API Key for pay-as-you-go billing.
  • Enter the project directory and start：cd /path/to/your/project then run step to enter the interactive TUI.
  • Describe the Task: In the TUI, directly describe requirements in natural language (such as analyzing project structure or modifying code); for script/CI scenarios, use step -p "任务描述" for non-interactive execution.
  • Initialize Project Guidance (Optional): Enter /init to generate AGENTS.md, an existing CLAUDE.md can take effect directly.
  • Manage Permissions: Use Shift+Tab to switch among the four permission levels: Ask / Read Only / Bypass / Autopilot. Dangerous commands are separately confirmed in any mode.
  • Resume Session：step -c to continue the most recent session,step -r to browse historical sessions and choose one to resume; within the TUI, you can use /resume。
  • Common Interactions：@ to reference project files,!命令 to directly execute Shell and hand it to the model,Esc Interrupt task,/hotkeys View all shortcuts.

Step Code's Core Advantages

  • Leading Benchmark Results: Terminal Bench 2.1 pass rate 80.9%, tied for first; on the self-built Multi-Frame long-horizon benchmark, 73.3% pass rate, ranking first.
  • Lowest Token Cost: Token consumption in both benchmarks is the lowest among peers in the same tier (long-horizon tasks average 5.09M tokens), saving more money at equivalent capability.
  • Reliable on Long-Horizon Tasks: Optimized for long-chain tasks in real-world development, with task completion stability superior to most frontier Agent Harnesses.
  • Efficient Context Management: By optimizing context management and the tool-calling chain, it reduces redundant information and ineffective calls, keeping Tokens more focused on advancing the task.
  • Built-in One-Click Publishing: The StepPage capability lets a static site go live with a single command after building, enabling a development-debugging-delivery loop within the terminal.

Step Code's Project Address

  • GitHub Repository：https://github.com/stepfun-ai/Step-Code

Comparison of Step Code with Similar Competing Products

Comparison Dimension Step Code DeepSeek-TUI
Tech Stack Not publicly emphasized (Agent Harness architecture) Written in Rust, keyboard-driven ratatui full-screen TUI
Open Source License MIT License MIT License
Supported Models StepFun Step series models Deeply bound to DeepSeek V4 (v4-pro/v4-flash), 1 million Token context
Benchmark Results Terminal Bench 2.1 pass rate 80.9%, tied for first; Multi-Frame long-horizon benchmark first No official public benchmark results
Token Cost Lowest in the same tier, long-horizon tasks average 5.09M tokens/task Auto mode automatically switches between Flash/Pro based on task complexity, dynamically controlling cost
Interaction Mode Four permission levels: Ask / Read Only / Bypass / Autopilot Three modes: Plan (read-only) / Agent (approval) / YOLO (fully automatic)
Featured Features Built-in StepPage one-click publishing of static sites LSP diagnostic integration, workspace rollback, sub-agent parallelism (up to 10), real-time cost tracking
Extensibility Open source and co-buildable, ecosystem just getting started Supports MCP protocol extensions, HTTP/SSE runtime API
Installation Method One-line curl script (macOS/Linux/WSL), Windows via WSL npm / Cargo / Homebrew / Scoop / Docker / Release binaries, native Windows support

Step Code's Application Scenarios

  • Getting Started with an Existing Project as a Beginner: For an unfamiliar codebase, directly use natural language to have Step Code analyze the project structure and sort out the logic, lowering the barrier to reading.
  • Feature Development and Bug Fixing: Describe requirements in the terminal (e.g., "add a captcha to the login API"), and it automatically completes code writing, debugging, and execution, suitable for improving daily development efficiency.
  • Long-Chain Refactoring Tasks: Use its long-horizon task capabilities and low token advantage to complete large-scale refactoring across multiple files and multiple steps, with controllable costs.
  • Automation scripts and CI integration: Via the step -p "任务描述" non-interactive mode, integrate into CI/CD pipelines or scheduled tasks to achieve scripted automated development.
  • Rapid static website delivery: After building pages locally, publish them online with a single StepPage command, suitable for rapid delivery scenarios such as personal homepages, documentation sites, and event pages.