StepFun open-sources Step Code for end-to-end development
StepFun released Step Code, an MIT-licensed coding agent that edits, tests and ships software from one CLI with strong long-horizon results.
What happened
StepFun open-sourced Step Code v0.1.0 on September 22 under the MIT License. The command-line agent is designed to cover the full software loop: reading a codebase, editing files, running tests, correcting failures and publishing a working result.
StepFun reports an 80.9% pass rate on Terminal-Bench 2.1, completing 72 of 89 tasks. It says that result tied for the highest score in its comparison while using fewer tokens than the other tied systems. On the company's 150-task Multi-Frame benchmark, Step Code completed 110 tasks, or 73.3%, while recording the lowest token use among six evaluated harnesses.
The release also includes StepPage, a built-in workflow for publishing a local static site to a shareable URL. Versioning and rollback are included, keeping development and delivery inside the same terminal environment. The tool runs on macOS, Linux and WSL, with Windows support still described as beta.
Why it matters
Step Code is notable because it treats the agent as a delivery system rather than a code-completion feature. The product boundary extends beyond generating patches: it includes execution, verification and deployment. That design reflects where coding agents are heading as users ask them to own longer tasks with less supervision.
The open-source license lowers the barrier for developers to inspect, modify and integrate the agent into private workflows. It also gives StepFun a way to compete internationally through tooling and community adoption even if its underlying model is not yet as widely distributed as those from OpenAI or Anthropic.
The evidence remains early. Step Code is version 0.1.0, its benchmark results come from StepFun's own evaluation, and Windows support is incomplete. The key question is whether outside developers can reproduce the reported reliability on messy repositories and production systems.