A new "official-stack" entrant in the terminal agent race
The terminal coding agent race in 2026 has moved past the question of who has one and into a contest of philosophies. Claude Code and Codex CLI fired the starting gun. MiniMax showed what an open-source, BYOK (bring your own key), multi-provider play looks like with minimax-code. And a clear trend has emerged: rather than waiting for the community to build shells around their models, model vendors are shipping the shells themselves. StepFun is the latest vendor to act on that logic. In late September 2026, the company open-sourced Step-Code, a terminal coding agent hosted at stepfun-ai/Step-Code. A GitHub API snapshot taken on September 27, 2026 shows 438 stars, 41 forks, TypeScript as the primary language, and an MIT license field, with the repository created on June 1, 2026 and the most recent push landing on September 26 — meaning the project was still under active iteration the day before the snapshot.
First, a note on how this piece relates to our existing coverage. We previously published a seven-dimension comparison of terminal coding agents, Terminal Coding Agent Comparison and Review, which scores MiniMax Code CLI, Claude Code, Codex CLI and others on a single rubric, plus a dedicated deep dive on MiniMax Code CLI. This article does not repeat that comparison. Instead it focuses on what makes Step-Code a distinctive data point: it is a new entrant, and it takes the "official model in an official shell" route — the exact opposite end of the spectrum from BYOK. Understanding the engineering trade-offs behind that route matters more than adding one more tool to your bookmarks.
Repository facts: cold numbers before the marketing
Start with the record. According to the GitHub API snapshot from September 27, 2026, stepfun-ai/Step-Code shows 438 stars, 41 forks, TypeScript, an MIT license field, a creation date of June 1, 2026, a latest push of September 26, 2026, and 44 open issues. One caveat on licensing: the GitHub API license field is what we report here, but the LICENSE file inside the repository remains the final authority — a basic habit when auditing any open-source project.
Now the positioning. The README opens with a sentence worth quoting verbatim: "Swift execution, long-horizon reliability, and high token efficiency." Those three phrases are not decoration; nearly every feature in the document serves one of them. The task definition is equally plain: Step Code runs in your terminal and handles the full task loop — reading code, making changes, and running tests. Compared with launches that lean heavily on narrative, this README is restrained: it talks about installation, sign-in, shortcuts, and migration, and almost never about vision. That restraint is itself a signal. The vendor is pushing Step-Code as an engineering tool, not a concept product.
One easily missed detail: the repository also ships a Simplified Chinese README, which drops the reading barrier for Chinese-speaking developers to nearly zero. For a Chinese vendor that is table stakes, but set against open-source projects that put up rough English docs first, the completeness still says something about the team's attitude.
Two roads: official model in an official shell, versus BYOK
The most interesting thing about Step-Code is not any single feature but its architectural bet. Per the README, the default entrypoint exposes exactly one built-in model provider: Step (StepFun). After sign-in, it discovers the models available on the active Step endpoint, and what you switch between in /model is simply the list Step returns. The shell is official, the model is official, and the two are tuned as a pair.
Concretely, that binding shows up in four profiles that differ only by service region and billing method. Step Plan CN and Step Plan Oversea both use browser OAuth with automatic credential refresh, and usage is included with Mini, Plus, Pro, and Max plans. Step Platform CN and Step Platform Oversea use API keys billed per request. Inside the TUI, /login opens the sign-in flow; from a shell you run step login. For API-key access you set STEP_API_KEY or pass --api-key on headless runs. Credentials land in ~/.stepcode/auth.json with mode 0600, and /logout clears them.
Place that next to MiniMax Code CLI and the contrast could not be sharper. BYOK's core promise is non-lock-in: your workflow is not married to one vendor's models; today you connect OpenAI, tomorrow Anthropic, and the agent itself never changes. The "official model in an official shell" promise is the mirror image: because the model and the shell come from the same family, token efficiency, tool-call formats, and long-task scheduling can be co-tuned in depth. You trade freedom for a higher experience ceiling.
Neither road is absolutely better; they match different situations. Teams allergic to vendor lock-in, or those that need to swap models freely, are safer on BYOK. Teams already using Step models, or those convinced by the claimed co-tuning gains, will find the integrated route less friction. This is the same judgment framework we apply in our AI Coding Agent Comparison: decide first whether you value freedom or optimization depth, then look at products.
Capability deep dive: /goal, StepPage, and ecosystem compatibility
Working through the README's feature list one item at a time.
Token efficiency and long tasks come first. The vendor says Step-Code is tuned alongside Step models to consume fewer tokens for the same task. Mechanically, long-horizon work is split across parallel subagents, each with its own isolated context, so redundant content never enters the main conversation. This is the same school of thought as the subagent architecture we discussed in Claude Code vs Cursor vs Codex: one coordinator directing a crew of executors, each keeping a clean working memory.
Then /goal. You hand an objective to /goal and Step Code works toward it autonomously, no hand-holding required. The companion /cron runs tasks on a schedule, and the status line shows a live timer for the active task. Taken together, these three details point in one direction: turning the agent from a question-answering tool into a resident executor. Long-running task delegation is one of the battlegrounds of 2026 agent products, and shipping both /goal and /cron in the core command set makes the product intent unmistakable.
Third is StepPage, the most underrated item on the list. The README describes it plainly: once your local page is ready, a single command publishes it as an accessible static website, with version management and rollback built in. Development, debugging, and delivery close their loop inside the same terminal. For prototyping, demoing to colleagues, or throwing up a temporary page, this removes the entire deployment step from the path.
Fourth is ecosystem compatibility, where the official team has clearly invested. MCP servers and Agent Skills work out of the box, and most Claude Code plugins are directly compatible, managed through /plugin. The migration details show real care: Claude Code and Codex MCP configurations are imported automatically on first launch, source files are never modified, the import is idempotent, and secrets are never inlined. An existing CLAUDE.md is used as-is, while /init generates the equivalent AGENTS.md project guide. Translated into plain language: the cost of moving over from Claude Code is, in theory, close to zero.
Fifth is the security boundary. Four permission modes — Ask, Read Only, Bypass, and Autopilot — cycled with Shift+Tab, and in every mode, dangerous commands require a separate confirmation dialog. Drawing that floor under the product is worth applauding.
Getting started: install, sign-in, and day one
Installation goes through the official installer, which downloads the latest release, verifies the checksum, installs step into ~/.stepcode/bin, and updates your PATH. On macOS, Linux, and WSL:
curl -fsSL https://static-openapi.stepfun.com/stepcode/install.sh | bashOn Windows, the PowerShell equivalent:
irm https://static-openapi.stepfun.com/stepcode/install.ps1 | iexNote that the README is explicit: Windows PowerShell support is currently in beta, and on Windows the recommendation is to install inside WSL. Verify in a fresh terminal with step --version and step --help, and upgrade later with step update.
After signing in, change into a project directory and launch:
cd /path/to/your/project
stepFor day one, the README offers two worked examples. Start with a prompt that lets the agent learn the project: "What is the tech stack of this project? What is each directory responsible for? How do I run it locally? Don't modify any files yet." — it reads without touching anything. Then give it real work, such as building a reference table for the error codes in src/api/errors.ts and saving it to docs/errors.md. The shallow-to-deep path is well designed: build trust first, delegate second.
Session management is standard: step -c continues the most recent session, step -r browses history, /resume finds past sessions inside the TUI, and /hotkeys lists every shortcut. If you prefer source, the repository builds with Git, Node.js, and pnpm: clone it, run pnpm install --ignore-scripts, pnpm run build, then pnpm step. Uninstallation is equally tidy: deleting ~/.stepcode removes all local data.
Benchmarks and cold second thoughts
The official news release (as relayed by an AI tools roundup on September 24) presents a striking set of numbers: a first-place tie on Terminal Bench 2.1 with an 80.9% pass rate and lower token consumption than peers, plus a leading 73.3% on Multi-Frame, the company's own long-horizon benchmark, with the lowest average token spend of 5.09M across six agent harnesses.
The framing matters: all three figures are vendor-reported by StepFun, not independently reproduced. Terminal Bench 2.1 is a public benchmark, so cross-vendor comparability is relatively strong, but the vendor still controls the test configuration and grading of its own submission. Multi-Frame is a self-built benchmark — questions, environment, and scoring all come from one party — which warrants a further discount. Sustained skepticism toward benchmark numbers is basic hygiene in this industry; we treat the broader problem in AI Agent Benchmark Trust: vendor-reported data is always worth reading, but decisions should not rest on it.
Now the cold second thoughts. For an open-source project less than four months old, 438 stars is a normal starting point — and that is all it is. At this size, the community ecosystem is unproven. Forty-four open issues is not alarming in absolute terms, but for a terminal agent, the rate at which issues accumulate and the quality of their handling will directly shape engineering maturity. More fundamentally, the success of the "official model in an official shell" route depends on the Step model family itself remaining competitive on coding tasks. A great shell cannot save a weak engine; conversely, if Step models keep improving, this deeply bound shell could become the best-feeling way to use them. It is a bet placed on the vendor's own models — worth watching, too early to score.
A pragmatic conclusion: teams already on Step Plan or Step Platform should try Step-Code immediately, since migration cost is near zero. Teams heavily invested in multi-provider setups should keep it on the watchlist until the ecosystem and independent reproductions mature. And the move common to everyone: treat the vendor-reported 80.9% as a lead, not an answer.
FAQ
Q1: Is Step-Code free and open source? A1: The repository is released under the MIT license (the LICENSE file in the repo is the final authority), so the code can be used, audited, and modified freely. Open source, however, does not mean free model usage: by default it only connects to the Step provider, and model usage is either included with Step Plan Mini, Plus, Pro, or Max plans or billed per request through a Step Platform API key.
Q2: Can Step-Code connect to OpenAI or Anthropic models? A2: Not according to the README. The default entrypoint exposes exactly one built-in provider, Step (StepFun); the four profiles only choose service region and billing method, they are not separate providers. If your team needs BYOK freedom across providers, the closer open-source counterpart is minimax-code.
Q3: How do Windows users install it? A3: An official PowerShell installer exists, but the README is explicit that this support is in beta, and on Windows it recommends installing inside WSL. The installer flags include --version for pinning a release and --install-dir for a custom directory.
Q4: I already use Claude Code — how painful is migration? A4: The README's migration path is smooth: MCP configuration imports automatically on first launch (source files untouched, idempotent, no inlined secrets), most Claude Code plugins work as-is and are managed with /plugin, an existing CLAUDE.md is used as-is, and /init generates the equivalent AGENTS.md.
Q5: Should I trust the 80.9% Terminal Bench 2.1 score? A5: It is a vendor-reported figure from StepFun, not an independent reproduction; the 73.3% on Multi-Frame comes from the company's own benchmark, with questions and grading from a single party. Treat it as a signal worth following, not a decision basis, and stay cautious about vendor benchmark claims.