FUXI · Smart Routing
Cost-aware · Auto-failover
Every request gets the best cost-performance
FUXI scores each request for complexity and automatically selects the most appropriate model tier. Exponential backoff auto-failover ensures resilience, while racing mode lets primary and backup models compete, using the fastest response to minimize latency.
FUXI · Thinking Loop
Like a human: think one step, act one step, think again
FUXI doesn't spit out answers all at once. It works like a human, in a loop: think, decide which tool to use, act, see the result, reflect on whether it's enough — if not solved, loop again with new findings. A hard problem may loop dozens of times, step by step, until it's truly done.
Each round:
think · act · observe · reflect
— if not solved, go around again, until it's genuinely right.
Quick Process
Repetitive tasks, clear steps — FUXI sees it and does it, no extra thinking. Renaming, formatting, simple file ops, instant results.
Simple stays simple — no waiting for AI to pretend it's thinking
Deep Reasoning
Real hard problems — FUXI pauses, breaks down the issue, reasons step by step, weighs options, then answers. Concurrency bugs, architecture design, edge cases — this layer handles them.
Real reasoning for hard problems, not guessing — the thought process laid bare
Verified Delivery
Thinking it through isn't enough — FUXI reads code, runs tests, confirms with real tools that the solution works, then delivers verified results.
Answers verified for real, not guessed — tests pass before delivery
FUXI · Persistent Sessions
Persistent memory, never lost
SelfPack Memory Engine
Goal Persistence
AI always remembers your project goals
Phase Tracking
Current progress and next steps clearly visible
Key Facts
Important decisions and data never lost
Cross-Session Continuation
Open terminal next day, seamlessly resume yesterday's work
FUXI · Multi-Window Collaboration
Three-Layer Collaboration, From Local to Cross-Machine
Multiple FUXI windows in the same project perceive each other in real-time, collaborate across machines through self-hosted relay servers, atomic task claiming prevents duplicate work.
mtime Hard Fallback Check timestamp before write, reject conflicts directly 丨UDS Soft Awareness Unix socket broadcast, same-machine awareness only
Local Multi-Window (UDS Broadcast)
Multiple FUXI windows in the same project broadcast file changes in real-time via Unix Domain Socket. Window A saves a file → all windows in the same project receive [peer-modified] → AI automatically re-reads context. The Peers panel on the right shows all active instances.
Dual-layer data security: mtime hard fallback + UDS soft awareness, zero-risk conflict writes
Cross-Machine Relay
Self-hosted fuxi relay-server, no cloud services required. Machine A in Beijing and Machine B in Shanghai exchange messages through the relay server, SendMessage bridge for cross-machine commands. Bearer token authentication, idempotent processing, built-in rate limiting.
Fully self-hosted, data never passes through any third party, Bearer token encrypted authentication
Persistent Task Distribution
Multiple windows share the same project-level task list (isolated by git root directory). When two windows try to claim the same task simultaneously, file locks serialize the claim action, only one succeeds. The Owner field records the claimant, preventing duplicate work.
Atomic file locks ensure unique claims, shared task lists automatically isolated by git root directory
A New Era
Starts Today
Download FUXI and run your AI coding agent in the terminal — no runtime, no complex setup, start coding instantly.
Get $15 usage on signup and start now
fuxi
