FuXi

FuXi is a terminal AI coding agent that reads code, edits files, runs commands and drives tools from a rich TUI. It is built in Go and ships as one static binary with no runtime dependencies. The current release is v0.1.8, and the source code is proprietary while the GitHub repository hosts documentation, installers and the issue tracker.

What It Is

FuXi is a provider-agnostic coding agent for the command line. You bring an OpenAI-compatible model or other supported provider, or sign in with FuXi-managed access, and FuXi adds an agentic Think → Act → Verify loop on top. The vendor positions it as an alternative to Claude Code. It runs entirely through a TUI with slash commands rather than CLI subcommands.

How the Workflow Runs

You launch fuxi, type a prompt such as fixing failing tests, and the agent works in your real codebase using 50+ built-in tools. These cover file read/write/edit, Bash and PowerShell, ripgrep search, web fetch with SSRF protection, LSP diagnostics, Jupyter, browser use, background tasks and parallel sub-agents. Shell commands pass through 27 rules and an AST safety classifier, and sensitive tool calls ask for approval unless a different permission mode is chosen.

Model Routing and Providers

Configuration lives in ~/.fuxi/config.yaml. Providers include Anthropic, AWS Bedrock, Google Vertex, Gemini, and OpenAI-compatible endpoints such as OpenAI, Azure OpenAI, DeepSeek, Grok, GLM, Qwen and Kimi. Optional routing tiers (fast, primary, reason) enable complexity-based routing, automatic failover and primary/backup racing.

Sessions, Memory and Extensibility

Sessions persist as JSONL transcripts with checkpoints, rewind, resume and forking. Idle "dreaming" consolidates memory across sessions, and context is auto-compacted. FuXi is an MCP client (stdio, HTTP, WebSocket, OAuth) and supports 27 lifecycle hooks, skills, plugins and custom slash commands, all hot-reloadable. The /team command starts coordinated parallel sub-agents.

Privacy and Evaluation

The README states that code and conversations are not collected by FuXi by default, and credentials stay in the local ~/.fuxi/ directory. The docs describe a small, self-run benchmark against another coding agent, state that it is not a third-party benchmark, and note that no scores on public benchmarks are published.