Fuxi Is a Free Terminal Coding Agent, But It Isn't Open Source

Fuxi installs as one static binary via a curl script, with no npm install, no Python environment, and no separate runtime to manage. Once it's running, every request you send gets scored for complexity and routed to a model tier automatically. Simple edits go to a cheap model, harder reasoning goes to a stronger one, and the tool switches between them without you picking a model yourself first. It's a young repo already past 1,100 GitHub stars two weeks after its first commit, and it's worth understanding both what it actually offers and one claim about it that doesn't hold up: despite how it's often described, Fuxi is not open source.

How the model routing works

Fuxi runs on what its own documentation calls a Think→Act→Verify loop: it reasons about what to do, takes an action (editing a file, running a shell command, searching code), checks the result, and repeats until the task is done. The routing decision sits underneath that loop. Every request is scored for complexity, and Fuxi picks a model tier to match, cheap models for simple tasks, stronger and pricier ones reserved for harder work, with automatic failover if a provider goes down and an optional mode that races a primary model against a fallback for latency-sensitive tasks. The README doesn't publish the actual scoring mechanism or thresholds, so you're trusting the router's judgment rather than auditing it, a real limitation for anyone who wants predictable cost control instead of automatic best-effort routing.

Fuxi is provider-agnostic for AI coding agent model routing. It works with OpenAI-compatible endpoints, Gemini, Bedrock and Vertex, and Anthropic's API, and you can either bring your own API keys for each provider or sign in through Fuxi's own OAuth flow for managed credentials. The OAuth path is the more interesting option if you don't want to juggle multiple provider accounts and billing dashboards, though it also means routing your usage through Fuxi's own account layer rather than directly to each provider.

A typical Fuxi session

Inside the terminal UI, you work through slash commands: /model to switch or inspect the active model, /config for settings, /cost to check what a session has spent so far. Beyond the Think→Act→Verify loop itself, Fuxi ships more than 50 built-in tools covering file operations, shell access, code search, LSP diagnostics, Jupyter notebook support, and MCP client support, so it can connect to the same MCP servers you'd already be running with Claude Code or another agent. Sessions persist to disk with checkpoint, resume, and rollback support, and an idle "dreaming" pass consolidates memory across sessions so long conversations auto-compact instead of eating your context window turn after turn. None of that is unique on its own, most serious terminal coding agents now do some version of session persistence and MCP support, but the combination with automatic cost-tiered routing is the part that stands out.

It's free, but it isn't open source

This is worth being precise about, because the label gets repeated loosely, including in some of today's own coverage of the project. Fuxi's binary is free to use for individuals, teams, and enterprises, no subscription required. But its LICENSE file states plainly: "Proprietary. All rights reserved," with no reproduction, distribution, or transmission of the source permitted without written consent from its maker, Shanghai YiTai Technology Co., Ltd. You get a compiled tool and its documentation. You don't get to read, audit, fork, or self-host the actual implementation. If open source specifically matters to you, whether for auditability, the ability to patch it yourself, or just principle, Fuxi doesn't qualify as an open source AI coding agent, even though it's free and behaves like an open project in every other visible way (public issue tracker, public docs, public install script).

That distinction matters more for a coding agent than for most tools, because a coding agent runs shell commands and edits your files with real permissions on your machine. Being unable to inspect what it's actually doing under the hood is a bigger trust ask than it would be for, say, a UI component library.

What's missing

Fuxi ships with no published benchmark scores. Its own README says as much directly: no SWE-bench, no Terminal-Bench, no Aider polyglot benchmark result, and it explicitly tells you to evaluate it against your own projects instead of trusting a leaderboard number. That's an honest stance, but it also means there's no third-party number to weigh against Claude Code, Codex, or Aider before you commit time to setup. At past 1,100 GitHub stars in its first two weeks, the traction is real and recent, but it's early traction, not proven performance.

Who Fuxi fits

Fuxi is worth trying if you want a lightweight, install-in-one-command terminal agent that automatically balances cost and capability across models without you manually switching, and you don't need to see or modify its source to trust it. It's a weaker fit if open source is a requirement for you, if you need a benchmarked track record before adopting a new tool, or if you'd rather keep full manual control over which model handles every request rather than delegating that decision to Fuxi's own router.

Sources: FuXi, GitHub