FuXi puts the loop and the routing wrapped around the model at the center of its pitch. Its README sums it up as "the model is the engine, FuXi is the vehicle": the configuration supports multiple providers and per-model settings, and the benchmark measures the agent loop rather than raw model scores.
The README pitches FuXi as a provider-agnostic alternative to Claude Code. It is built in Go, ships as one static binary with no runtime dependencies, and asks you to bring whatever OpenAI-compatible model you already use. On top of that model it runs a Think, Act, Verify loop, with cost-aware routing across LLM providers and automatic failover.
Providers as a catalog
The config format shows how far the swappable-engine idea goes. A single-provider setup is four lines of YAML: provider, base URL, API key, model. For more than one, FuXi offers a layered schema, with a providers: catalog where each entry declares its type, endpoint, key, and models, and a separate model: layer that selects the active provider and model id. The README says the layered schema supports multiple providers and per-model settings.
The README separately claims cost-aware routing across providers and automatic failover. The README also lists a fuxi proxy subcommand that starts a smart routing proxy doing protocol bridging between providers, and fuxi launch, which runs a proxied binary through that proxy using your FuXi config. What the README does not spell out is the routing policy itself, so if you care how it picks a model for a given request, that is the part to test.
Credentials follow the same pattern. You can supply a provider key (the README names OpenAPI-compatible endpoints, Gemini, and Bedrock/Vertex), export FUXI_API_KEY, or run fuxi login, which provisions FuXi-managed models with no key needed. fuxi init writes a starter config and auto-detects a provider from environment variables already set, and fuxi wizard walks through provider, base URL, key, model, and a connection test. For a single run, -m, -P, -b, and -k override the model, provider type, base URL, and key.
The loop carries the weight
If the model is interchangeable, the loop has to do the work. The README claims the Think, Act, Verify loop and its routing let a model perform above its raw benchmark. Behind the loop sit 50+ built-in tools in the same binary: file read, write, and edit, bash or PowerShell, ripgrep search, web fetch, LSP diagnostics, Jupyter, browser use, background tasks, and parallel sub-agents.
Autonomy comes with controls attached. Shell commands pass an AST safety classifier before they execute, with fine-grained permissions and audit logging on top. An auto flag approves safe tool calls on its own, gated by that classifier plus a circuit breaker, and the permission-mode flag accepts default, plan, or bypassPermissions. Transcripts persist to disk, checkpoints let you resume, roll back, or fork a session, and long conversations auto-compact to save tokens. The README also describes an idle "dreaming" step that consolidates memory across sessions.
Reading the benchmark in context
The central claim rests on a head-to-head against Claude Code, and the README explicitly identifies its limits. Both agents ran through their own native clients on identical baselines, scored by pytest plus coverage, across 15 micro dimensions and 4 large-project dimensions, with methodology, raw results, and exact commands kept in the repo. The same section describes it as a small, self-run task set rather than a third-party benchmark, notes that it measures the agent loop and not raw model scores, and asks readers to treat it as a data point. FuXi currently has no published score on SWE-bench, Terminal-Bench, or the Aider polyglot benchmark.
In place of headline numbers, the README hands you a checklist. Run fuxi doctor to check config, API key, git, and ripgrep, then fuxi verify to confirm the provider connection. Point FuXi at a failing test in one of your own projects, then push the identical task, model, and context through another tool and compare correctness, tool coverage, cost, and iteration time. The README says /cost, /usage, /context, and /status expose what you need for that comparison inside the TUI.
One licensing detail before you adopt it: the README's license badge reads Proprietary, while the highlights describe the binary as free forever, with no license cost for individuals, teams, or enterprises. No license cost is a different claim from open source.
GitHub: https://github.com/fuxicodex/Fuxi
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)