Terminal AI agent that switches LLM tiers mid-task to keep costs down

What it does

FuXi is a single static binary that drops a persistent AI coding agent into your terminal. It reads and edits files, runs shell commands, searches code, and delegates work to sub-agents, all from a text UI. The agent works in a loop—reason, act, verify—persisting session state to disk so you can resume, fork, or roll back conversations.

The interesting bit

The standout is its cost-aware router: every request is scored for complexity and sent to the cheapest suitable model tier, with automatic failover if a provider flakes. That sounds like a minor billing optimization, but it is really an admission that most agent work is boilerplate that does not need frontier-level inference.

Key highlights

• One static binary, no runtime dependencies; self-updating with checksum verification.
• 50+ built-in tools including shell, ripgrep, LSP diagnostics, Jupyter, browser use via MCP, and parallel sub-agents.
• Safety guardrails via AST classification and fine-grained permission rules before any shell execution.
• Durable sessions with checkpoints, rollback, and an idle “dreaming” pass that consolidates memory across conversations.
• Bring-your-own-key or OAuth sign-in; acts as an MCP client and can itself run as an MCP server.

Caveats

• The license badge reads “Proprietary,” even though the README claims the tool is “free forever” for individuals and teams—so the exact freedom to modify or redistribute is unclear.
• It currently ships without published scores on third-party benchmarks (SWE-bench, Aider polyglot, etc.); the authors invite users to evaluate it themselves on real tasks.

Verdict

Developers who want a self-hosted, terminal-first alternative to cloud-centric coding agents should look here—especially if you are tired of burning API credits on simple refactors. Those seeking an open-source codebase or published benchmark leaderboard will not find either.