Intel / article

StepFun open-sources Step Code, a Claude Code-compatible CLI agent

Source: github.com
Author: StepFun
Date: Published Sep 22, 2026

Context

StepFun released Step Code v0.1.0, an MIT-licensed coding that runs in the terminal and, according to its README, reads code, makes changes and runs tests. It is built around StepFun's own models: the default build exposes one built-in model provider, Step, reached through a Step Plan subscription or a pay-per-request Step Platform API key, and lists the Step models available to that account.

The Claude Code comparison comes from its migration features rather than its models. On first launch Step Code imports existing Claude Code and Codex configuration without modifying the original files, uses an existing CLAUDE.md project guide as-is, and says most Claude Code plugins work unchanged. MCP servers are the connectors coding agents use to reach outside tools and data, so carrying them over lets a developer try Step Code without rebuilding that setup. The work itself still runs on Step models.

StepFun also built in /goal, which hands the agent an objective to pursue on its own, /cron for scheduled runs, parallel subagents with separate contexts, which it says keep redundant content out of the main conversation and reduce use, and StepPage, which publishes a local directory as a static website in one command. StepFun's announcement reports 80.9% on Terminal-Bench 2.1, described as in , and 73.3% on Multi-Frame, a 150-task long-horizon benchmark StepFun built itself. Neither the post's text nor the README says which Step model or settings produced those scores, so they are StepFun's own results rather than an independent comparison with other coding agents.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source github.com