Muxee

NVIDIA Switchyard, productised

One key.
Every model.
Routed right.

Point your coding agent at Muxee with a two-line config change. From then on every call it makes is routed to the most cost-effective model that can actually do the job — frontier models for the deep reasoning and the recovery, efficient models for everything else. Your agent never notices.

Start free trialSee how it works

14 days free, no card. Bring your own provider keys or use ours.

[object Object]Requests leave your coding agent, pass through the Muxee router, and are sent to whichever model can do the job most cheaply. Most land on efficient and mid-tier models; the expensive frontier model is used only for the hard turns.Your agentone key, one URLMuxeerouterFrontierdeep reasoning, recovery$$$Mid-tiermost real work$$Efficient / localmechanical steps$Saved vs frontier-only$1,284
Illustration, not live data. Your own saved-vs-frontier-only figure is computed from your real traffic and shown in the dashboard.

Two lines of config. Nothing else changes.

No TOML, no Rust binary, no server to run. Muxee runs the routing engine so you point at a URL.

  1. 01

    Create a router

    Pick a preset — Auto is the recommended default — choose which models it may use, and Muxee gives it a stable model ID like muxee/auto.

  2. 02

    Point your agent at Muxee

    Set the base URL and your Muxee API key in the agent you already use. Two lines, once. The dashboard generates the exact snippet for Claude Code, Codex, OpenClaw or a plain SDK.

  3. 03

    Work normally

    Every call your agent makes is routed per turn. You get the selected model and the rationale back on every response, and the savings against a frontier-only baseline in the dashboard.

~/.zshrcClaude Code
# point Claude Code at Muxee
export ANTHROPIC_BASE_URL=https://api.muxee.app
export ANTHROPIC_AUTH_TOKEN=mx-live-7f3c2a91e4b8…d0a4
export ANTHROPIC_MODEL=muxee/auto

Example key. Yours is generated in the dashboard and shown once.

Seven routing strategies, tuned for you

Start with Auto. Change your mind later — a router is a preset plus the models you let it use, and every knob underneath is yours if you want it.

  • AutoRecommended
    Reads your agent's own signals. Free routing decisions.

    Starts efficient and hands over to the capable model the moment your agent signals it is struggling. No judge call, so the routing itself costs nothing.

    muxee/auto

  • Economy
    Max savings. Runs cheap, escalates only when the run struggles.

    Every turn runs on the cheap model. A judge reads the finished turn and escalates only after repeated struggle signals.

    muxee/economy

  • Guarded
    Predicts difficulty up front. Fails toward quality.

    A judge predicts how hard the turn is before it runs, and anything above the threshold goes to the strong model. Errs toward quality.

    muxee/guarded

  • Domain
    Your own rules: coding here, math there. Schema-driven.

    You write the prompt and the JSON schema; the judge picks a target by name. For teams who already know which model owns which kind of work.

    muxee/domain

  • Split test
    A/B two models on live traffic; prove the difference.

    Weighted, content-blind split across two or more models. Fix the seed to make the split reproducible.

    muxee/split-test

  • Direct
    One model, Muxee metering and analytics.

    No routing at all. Useful as a baseline to measure the other routers against, and for work you never want moved.

    muxee/direct

  • Smoke
    Free pipeline test. No upstream call, no tokens.

    Answers "OK" without calling any model. Point your agent at it once to prove the key, the base URL and the router all work.

    muxee/smoke

The routing works. Here is who measured it.

Muxee does not invent a routing algorithm. It runs NVIDIA’s NeMo Switchyard, which independent teams have already benchmarked.

74%1

lower cost than a frontier-only baseline

Escalation routing between Nemotron 3.5 Lightning and Claude Opus 4.8 sent just 7% of calls to the frontier model, at a measured ~6-point accuracy tradeoff.

Muxee preset: Economy

2.8 pts2

of frontier accuracy, at ~28% lower mean cost

Staged routing between Opus 5 and Kimi K2.7 scored 50.6% at a $3.11 mean cost — within 2.8 percentage points of Opus 5 alone.

Muxee preset: Auto

These are published results from other teams on their own benchmark suites, not a promise about your traffic. Your saving depends on your models, your prompts and how much of your work is genuinely hard. Muxee shows your real number — measured against a frontier-only counterfactual on your own usage — in the dashboard from your first request.

  1. 1LangChain, Switchyard agent-routing benchmark: 145 multi-turn agentic tasks from its internal deep-agents evaluation suite, averaged over five runs. Muxee’s Economy preset uses this routing strategy.
  2. 2Cognition, staged routing in Devin Desktop measured on FrontierCode Main, Cognition’s production-grade coding benchmark. Muxee’s Auto preset uses this routing strategy.

What would it be worth on your bill?

Drag the two numbers you already know. The band is deliberately wide — only your own traffic can narrow it.

Monthly LLM spend today$400

$50 to $10,000+

Share of that you would route through Muxee70%

Work you keep pinned to one model stays pinned — only routed spend can be saved against.

Estimated monthly saving

$84 – $196

$1,008 – $2,352 a year, on $280 of routed spend.

An estimate, not a quote. It assumes 30% to 70% off the spend you route — a band chosen to sit at or below published routing benchmarks, not at the top of them. Your real figure depends on your traffic, and Muxee measures it against a frontier-only counterfactual from your first request.

Find out on your own traffic

Questions worth asking

Latency, privacy, your own keys, and what happens when something upstream breaks.

Route your first request in five minutes

Create a router, copy two lines into the agent you already use, and watch the decisions land in the dashboard.

Start free trialSee pricing