Duel

The decentralized router for LLMs.

Route everything through duel-auto, cheaper & censorship-resistant.

[Duel @useDuel]

Picking the right LLM is a full time job.

What if every model competed for your money in real time and you only paid for the winner?

https://t.co/0LvEw2mDRE

[Duel @useDuel]

Duel Agents has been listed on @crestingdev 👋

https://t.co/F9eGTjkZs7

[Duel @useDuel]

2 things are happening at once: every new frontier model launches smarter and more expensive.

Ramp's data shows the top 10% of companies already pay $73k+/month in tokens, the top 1% over $800k/month. Goldman Sachs says AI spend is approaching 10% of total headcount cost at some firms and could rival it within a few quarters.

That's the exact problem Duel exists to solve. You stop picking a model and start routing, the cheap model handles what doesn't need the expensive one, the expensive one only gets called in when it's actually needed while sub-agents and mini-agents each handle different branches of the tree.

The winners of the next few years of enterprise AI won't be whoever has the best model, it'll be whoever routes across all of them, that's the bet behind $DUEL.

[Aaron @itsaaronshaw]

Stop overpaying for LLM calls.

Here's how to route every prompt through the cheapest model that still wins in about 2 minutes, without rewriting your app.

Duel is an IDE-native routing layer. It's OpenAI + Anthropic wire-compatible, so your existing code, SDKs and agents just work (you only change the base URL)

Get a key Sign up at https://t.co/X1FCrmXfgs → subscribe → create an API key in the dashboard.

export DUEL_API_KEY=duel_yourprefix_yoursecret

Install for your tools
npx @duel-agents/install all
npx @duel-agents/install doctor # verify
Works with Claude Code, Cursor, Codex CLI and OpenClaw out of the box.

Or build on it directly

from langchain_openai import ChatOpenAI
llm = ChatOpenAI(
model="duel-auto",
base_url="https://t.co/g2CcA9Lgol",
api_key="duel_yourprefix_yoursecret",
)

duel-auto = let the router pick. Same for the TS SDK, LangChain and LlamaIndex (pip install langchain-duel).

Star it + full docs 👇 https://t.co/TWsHKnM3tr

[Duel @useDuel]

What we're building at Duel:

- Routing every prompt across a decentralized pool, picking the cheapest model that clears the task instead of defaulting to one provider's flagship

- Splitting bigger jobs into sub-agents and mini-agents, each routed independently so you're billed like 1 intelligent system, not a room full of models guessing

- Provider-agnostic access that survives any single model being restricted, deprecated or taken down so no one gatekeeper can narrow what your AI can do

- Open and self-hostable models in the pool, with a path to verifiable routing: proof of where your prompt went, not just our word for it

Decentralization you can verify, not a trust-me-bro in the terms of service.

[Duel @useDuel]

There's a quiet shift happening in AI and teams are catching up to it.

When 1 company hosts the model you use, they hold your prompts, decide what it's allowed to answer and can change the price or pull access overnight.

Your privacy, your cost and what you're even permitted to run all become someone else's policy decision.

It's why @AskVenice is building private, uncensored inference, and why networks like @bittensor are pushing AI onto infrastructure no single party controls.

We think that shift has to reach the routing layer too

Here's where Duel is taking it:

[Duel @useDuel]

One prompt in, a full execution tree out.

Routing at the top. Sub-agents for structure. Mini-agents for speed.

All of it running across decentralized inference, with no single provider deciding what you're allowed to run. Billed like you ran one intelligent system, not a room full of models guessing at a job they were never built to carry alone.

This is where the real efficiency starts.

[Duel @useDuel]

Routing was step one: pick the model that actually wins the task, chosen across a decentralized network instead of one provider's lineup.

Step two is what happens after. Some jobs are too large, too layered, too expensive to force through a single model in one flat pass.

Once a winner clears, the work splits.

Sub-agents take the branches, mini-agents take the leaves.

[Duel @useDuel]

Sub-agents break the job into parallel workstreams. Every piece runs the same logic: set the field, eval picks the winner, cheapest output that still solves its part.

Mini-agents sit below them and handle the smallest units of work, on the lowest tier that clears.

Same standard at every depth. No agent drowning in context, no model burning premium tokens on work that should have been divided three layers ago.

[Aaron @itsaaronshaw]

1,000 stars on duel 🙏

Duel was basically a side project with a dozen testers I knew by name, now there's an actual community around it, ppl sending PRs i never asked for, building integrations into tools i hadn't even considered..

Honestly a good chunk of what we're shipping next comes straight from community recs, Duel is the fastest thing i've ever worked on and we're barely getting started.

https://t.co/TWsHKnM3tr

[Aaron @itsaaronshaw]

Number wise.. 🥁 we're at about $10k MRR and growing quick, I haven't spent a cent on ads to get here, it's all word of mouth, 1 dev plugs it in and tells the next one.

GTM strategy for next month: get our first enterprise contract and 10x the volume, our infra's built for it.

Grab your Duel API key at: https://t.co/X1FCrmXfgs

[Duel @useDuel]

Duel routes your model calls across providers, including open and decentralized ones so when one closes a door, your app keeps running.

As of today, that routing is native in @LangChain and @llama_index the 2 frameworks most teams already build on.

pip install langchain-duel
pip install llama-index-llms-duel

https://t.co/TiUGbfIcGu https://t.co/PQb4QzMwFb

PR: https://t.co/78fjJCAsjK

[Duel @useDuel]

Centralized AI inference is starting to show its cost.

Fable 5 was restricted, OpenAI is moving toward requiring ID to use its API, a handful of providers now decide who gets to run what and when.

Duel exists so you don't have to depend on any single one of them. Run inference without asking permission, handing over ID or betting your product on one company's terms staying the same.

[Aaron @itsaaronshaw]

Sending every prompt to GPT-5 or Opus is the most expensive mistake in AI right now.

1 request is never one task, it's a dozen.

A couple of parts need a genius while the rest just need someone who can follow instructions but you pay genius rates on all 12, every single time and your bill quietly doubles for work a cheap model would nail.

@useDuel fixes that at the root, we break every prompt into sub-agents and mini-agents, so the hard parts get a heavy model and everything else gets a cheap one.

Live for 1 week and already at 1k+ GitHub stars and you can see what you save in real time👇

[Aaron @itsaaronshaw]

The simplest way to say it: @useDuel is the decentralized AWS for intelligence.

AWS killed the idea that you buy one giant server and push everything through it, you rent exactly the compute each job needs and pay for nothing more.

We do that to models. Every sub-agent and mini-agent runs on the smallest model that can actually finish its piece, in parallel then stitched back into one answer.

Our decentralized router is ho w teams average 73% lower bills without giving up quality.

Nobody pays a flagship price for every keystroke anymore. The one-model-for-everything era is already dead, we're what comes next.

API key at https://t.co/X1FCrmXfgs.

[Marco Ramilli @Marco_Ramilli]

🤖 Duel-Agents
⭐ 739 stars

Stop burning cash on LLMs. Run prompts against multiple models and auto-pick the cheapest response that still wins, straight from your IDE.

🔗 https://t.co/M9DQ5pa02B

#AI #MachineLearning

[Duel @useDuel]

@Marco_Ramilli 👍

[Duel @useDuel]

Early testing: ↓ 73% API spend*

Most "routing" today is guesswork dressed up as infrastructure. It picks a model before knowing if it actually fits the task and you pay either way.

We do the opposite. We use public benchmarks, academic research and real model studies to know which models are good at what, then send each prompt to the one most likely to win it and keep your prompts private while we do.

*early data, will vary by workload

[Duel @useDuel]

Static LLM routing is already dead. It just hasn't been replaced yet.

The old way: you pick one model up front and send everything to it. When it's the wrong tool for the task, you get a worse answer and still pay full price.

The new way: every task becomes a live market. Models compete on the actual work, across a decentralized pool instead of one provider seeing everything, and you keep the answer that wins.

That's the layer we're building at @useDuel

[Duel @useDuel]

.@hackernoon wrote an article about Duel.

The main idea: the model you use for a task shouldn't be hardcoded. It should be picked per prompt, at runtime, across a decentralized pool instead of routing everything through 1 provider.

If you lock in on a flagship model, every new release means redoing your setup and overpaying for every small task. With routing, a new model just joins the pool, your prompts stay private and each task goes to whichever model handles it best at the cheapest price.

https://t.co/CBELNYDBOv