Moamao

One API call. A panel of models behind it.

Moamao turns a single prompt into a structured deliberation — parallel drafts, cross-model critique, judged synthesis — and returns one answer with a trace you can audit. Built for people already paying for Claude and GPT who want them working together instead of pasted into separate tabs. Here is each piece in detail.

The pipeline: three stages, in order.

A run is a real sequence: the output of each stage is the input to the next. The numbering below is the order your prompt travels.

  1. 01

    Propose

    "panel": "balanced-4" // per-seat control arrives in beta

    Your prompt fans out to the panel — up to four proposer models drawn from different provider families (Claude, GPT, Gemini, DeepSeek, and more). Each answers independently and in parallel, with no visibility into the others. Independence matters here: correlated drafts would give the later layers nothing to work with. Because the fan-out is parallel, layer one takes roughly as long as the slowest single model, not four calls in sequence.

    Panel makeup is yours to set — pick a preset today; naming exact models per seat arrives through the beta. Cross-provider panels are the default because models trained on different data fail in different places, and because the panel is yours to configure, not fixed by us.

  2. 02

    Critique

    "layers": 2 // propose → critique

    Every proposer now reads the other drafts. Each model flags claims it disagrees with, points out gaps, and revises its own answer in light of what the rest of the panel wrote. This is where a single model’s hallucination tends to die: an invented citation or a wrong API signature rarely survives review by models that didn’t make the same mistake.

    The critique layer also runs in parallel — the revised drafts come back in one round trip. You can add layers for harder problems or drop to one for speed; two is the default.

  3. 03

    Aggregate

    "aggregator": "auto" // or pin a specific judge

    A judge model reads the revised drafts and synthesizes one final answer — not a vote, not a concatenation, but a new response built from the strongest parts of each draft. Alongside the answer it records disagreement notes: the places where the panel didn’t converge, so you can see exactly where confidence is thin.

    Set the aggregator explicitly or leave it on auto and Moamao picks a judge suited to the task. The judge streams as it writes, so the final answer arrives like a chat response.

This is the step the numbers back up: a mixture of models has been shown to beat its own best single member, 57.5% → 65.1% on AlpacaEval 2.0 (LC win rate) (Wang et al., 2024 — Mixture-of-Agents). That is the lift the judge is built to capture.

Smart routing: not every prompt deserves a panel.

Before a run starts, Moamao scores the prompt. Simple, factual, or low-stakes requests — a date format, a one-line rewrite, a lookup — skip the panel entirely and go to a single fast model. You get the answer in one model’s latency instead of two sequential layers, and you pay a fraction of a run instead of a full mixture.

Anything ambiguous, high-stakes, or genuinely hard goes to the full panel. The routing decision is recorded in the trace, so you can always see why a request went where it did — and you can force the panel on any request if you disagree with the router.

In practice this is what makes mixture economics work: the panel’s cost and latency are spent only on the prompts where four models beat one.

Bring your own keys: your provider accounts do the model work. We orchestrate.

BYOK is the default and the honest way to run a mixture today: you connect your Anthropic, OpenAI, Google, and open-weights provider keys, and every model call runs on your accounts, your rates, and your rate limits.

How keys are stored

Keys are encrypted at rest, held per provider, and used only to place the calls your runs require. They are never logged, never shared between customers, and never used for anything resembling training.

What Moamao charges

An orchestration fee per run — that’s it. The model usage itself lands on your provider invoices at whatever rates you negotiated. Two bills, both readable: your providers meter tokens, we meter runs.

Moamao Cloud — live

Don’t want to manage four provider accounts? Moamao Cloud is our managed mixture endpoint — it runs the full mixture on managed capacity with zero provider setup, metered per token. Paid plans include a monthly Moamao Cloud credit equal to their price; even Free includes a small one.

Traces and observability: every run leaves a record you can interrogate.

A mixture you can’t inspect is just a slower black box. Moamao stores the full deliberation so the answer is never the only artifact.

Replay any run

Every draft, critique, and judge decision is stored per run. Open a trace and step through the full sequence — what each model said in layer one, how it revised in layer two, and what the judge kept.

Per-model attribution

The trace marks which panel member contributed which part of the final answer. When an answer is wrong, you can see whether one model misled the judge or the whole panel agreed on the error.

Disagreement notes

The judge records where the panel split and how it resolved the split. High-disagreement runs are your signal to double-check — the seams are visible instead of smoothed over.

Retention you control

Traces are encrypted at rest. Moamao Cloud runs help train and improve our own models — you can opt out anytime, and BYOK runs never do. Keep them for auditing, or turn on zero-retention mode (Scale plans and above) and Moamao discards run content after the response is delivered and doesn't use it for training.

Panel presets: start from a tuned panel.

Presets are starting points, not walls — every field is overridable per request, and custom panels are available on paid plans.

code-review

Panel makeup
Two strong coding models + two generalists, 2 layers, code-tuned judge
Best for
Reviewing diffs, migration plans, API designs — anywhere a missed edge case is expensive

research

Panel makeup
Four models from four different provider families, 2 layers, citation-strict judge
Best for
Fact-heavy questions where cross-checking claims matters more than speed

longform

Panel makeup
Three writing-strong models, 2 layers, style-preserving judge
Best for
Drafts, documentation, and editing passes that need one coherent voice at the end

balanced-4

Panel makeup
Four models, each from a different provider family, 2 layers, auto judge
Best for
The default. General-purpose prompts where you want maximum viewpoint diversity

What a run costs: one line on the invoice.

With BYOK, model usage stays on your provider bills and Moamao charges a flat orchestration fee per run. With Moamao Cloud, usage is metered per token across the whole mixture. Either way it lands as one line on one Moamao invoice. Routed-to-single-model requests count as a quarter run. Failed runs don’t count at all. Beta plans start at $0. You don’t need a bigger model or your own inference cluster to get a stronger answer: combining frontier models you already have access to gets you there.

See beta pricing
terminal
POST https://api.moamao.com/v1/runs
Authorization: Bearer $MOAMAO_KEY

{
  "panel": "balanced-4",
  "input": "...",
  "layers": 2,
  "aggregator": "auto"
}

// billed: 1 run
// (0.25 if routed to a single model)

Open beta: stop betting everything on one model.

Create a free account, generate an API key, and run your first mixture in minutes — or connect a repo in Build and watch it work. BYOK from day one.

Create a free accountOpen Build