Documentation · updated October 2026

How Dyad works

Everything behind the app: how a pair is chosen, what each stage is asked to do, how credits are counted and bought, how agents work, what we store (almost nothing) and how to call it over HTTP.

01Overview

Dyad is a chat workspace built around one idea: a single model should not be the only reviewer of its own answer. In Pair mode every question passes through two models from different providers — one writes, the other audits — and the final text is written with both in view.

The product has four surfaces:

  • Pair — draft, challenge, resolve. For questions where being wrong is costly.
  • Chat — one model, streamed, fastest and cheapest.
  • Duel — two models answer blind, you vote, then see who wrote what.
  • Image — text-to-image with a local gallery.

Around them sit credits (how usage is paid), Post & Earn (how usage can be earned), and $DYAD (the participation token, separate from credits).

Free tier: until the provider account is funded, every Dyad name runs on a free open model (the trace shows which), Apex and Summit stay locked, and images are rendered by a free Flux model. If a free model is busy, Dyad switches to the next one within a few seconds.

02Pair mode

Draft · AComplete answer, committed, no hedging.
Challenge · BAudit only: ✕ problems, ✓ checks, verdict.
Resolve · BFinal answer written from draft + audit.

Draft

The drafter gets your message, recent conversation, attachments and optional web results. It is told to commit to a concrete answer and not to mention that it will be reviewed — hedged drafts are hard to audit.

Challenge

The challenger never sees the drafter's instructions, only your request and the finished draft. It is told not to rewrite, only to audit: factual errors, unsupported claims, hidden assumptions, missing cases, logic gaps, security issues and better alternatives, capped at eight points. Temperature is lowered to 0.3 for this stage.

Resolve

The challenger then writes the final answer from your request, the draft and its own review. It keeps what the review did not refute, applies what it did, and discards review points that turn out to be wrong. It only mentions the process if a real disagreement is left open.

Skipping the challenge

Settings → Challenge stage off sends the draft straight to resolution. You pay for two stages instead of three and wait less, but there is no independent audit. The answer's trace records that the review was skipped.

Rule of thumb: a pair costs about 2–2.5× the time of a single answer. Use Chat for lookups and quick rewrites; use Pair for code, plans, math, and anything you would otherwise double-check yourself.

03Auto Pair and the roster

With either picker set to Auto, Dyad reads the prompt and chooses a pair. Hard rules come first (an image forces a vision drafter; drafter and challenger must differ), then a category match:

SignalDrafterChallenger
Image attachedLensPrism
Code, APIs, errorsForgeMeridian
Math, proofs, calculationsAtlasField
Long or analytical promptAtlasOrbit
Everything elseFieldMeridian

Pairs deliberately cross provider families: two models trained on similar data tend to share blind spots, which makes the review weaker.

Names vs engines

Dyad names (Spark, Atlas, Summit…) are stable. The engine behind a name can be upgraded; every answer's trace records the exact engine used, so older answers remain attributable. The live roster is at /#models and GET /api/models.

04Chat, files and web

Chat mode sends your message to one model with the same history, file and credit handling as Pair mode.

Attachments

  • Images — sent to vision-capable models (marked ◐). If you pick a text-only model, Dyad switches to Lens for that turn.
  • PDF — parsed to text by the provider's document parser before the draft.
  • Text & code (.txt, .md, .csv, .json, .py, .js …) — inlined into the prompt, up to 60,000 characters each.

Limits: 6 files, 3 MB each, 4 MB per request. Attachments are sent only with the turn they belong to; later turns see the conversation text.

Web

The web toggle lets the drafter pull up to three live search results. It adds provider cost, which is reflected in the credits for that stage. The challenger and resolver work from the draft, not from fresh searches.

Controls

Temperature (0–1.2), answer length (short / medium / long caps the tokens per stage), stop (Esc or the Stop button), retry, copy, export to Markdown, and incognito threads that are never written to storage.

05Image studio

Image mode sends a prompt and an aspect ratio (1:1, 16:9, 9:16, 4:3, 3:4) to the Canvas model and returns a PNG. Generations are saved to your browser's IndexedDB, so the gallery survives reloads without leaving your device. Each image costs about 15 credits.

06Agents

An agent is a saved setup: a name, an icon, standing instructions, a mode (Pair or Chat), the models it uses and a few starter prompts. Pick one in the sidebar and every message in that chat follows its instructions.

Built-in agents

AgentSetupUse for
Launch StrategistPair · Atlas → OrbitToken launch plans: liquidity, timing, community, day-by-day checklists
Tokenomics AuditorPair · Atlas → MeridianSupply, vesting, taxes and allocation, with the risks holders will see
Contract ReviewerPair · Forge → MeridianSmart-contract review: permissions, mint and tax switches, upgrade paths
CT CopywriterChat · OrbitPosts and threads for X in a builder voice
Degen Risk CheckPair · Field → PrismRed flags before you buy: holders, liquidity, deployer history

Your own agents

+ new in the sidebar opens the editor. Built-in agents can be duplicated and changed. Agents are stored in your browser like your chats.

Sharing an agent

Copy share link in the editor packs the whole agent into the link (after #agent=). Whoever opens it gets the agent added to their own sidebar; nothing is uploaded.

07Credits

Credits are the only thing that pays for model usage. They are not the token, not tradeable, and not an investment.

credits = provider_cost_usd × 3000 (3 credits per $0.001 of compute)

When a provider does not report cost, Dyad estimates from tokens and the model's tier:

TierCredits per 1k tokensModels
Light1Spark, Swift
Standard2.5Field, Atlas, Forge, Lens
Advanced6Meridian, Orbit, Prism
Frontier20Apex, Summit

Each stage is charged separately after it finishes, so a stopped or skipped stage is never billed. The per-stage breakdown is in every answer's trace. New accounts start with 250 credits; more can be earned through Post & Earn and, after launch, from the staking credit pool.

08Buying credits

Credits can be bought in the app (Buy credits in the sidebar, or the wallet button in the top bar) and are paid on Solana, in USDC or in $DYAD at 20% off once the token is live.

PackPriceCredits
Spark$54,000
Orbit$1513,000
Nova$4038,000

How a purchase works

1 · QuoteThe server signs a price quote for your wallet, valid 15 minutes.
2 · TransferYour wallet sends the exact amount to the Dyad treasury.
3 · VerifyThe server checks the transaction on-chain, then the credits are added.

No account and no email: the transaction is the receipt. Receipts with explorer links are listed in the store. If you close the tab before the check finishes, reopening the store resumes it.

Wallets

  • Desktop — Phantom, Solflare or Backpack browser extension. The site never asks for your wallet on page load; it connects only when you press Connect wallet.
  • Phone — mobile browsers have no wallet inside them. Press Connect wallet and choose Open in Phantom or Open in Solflare: Dyad opens in the wallet's own browser with the store ready.
The first purchase to a fresh treasury token account costs about 0.002 SOL extra in network rent; after that only the normal Solana fee applies.

09Privacy

  • History is local. Conversations, settings, credits and generated images live in your browser (localStorage and IndexedDB). There is no chat database on our side.
  • The server is a relay. It forwards the current request to the model provider, streams the answer back and keeps nothing. It never logs prompt contents.
  • Providers receive the request needed to answer it, under their own API terms.
  • Incognito threads are not written to storage at all.
  • Consequence: clearing site data deletes your history permanently. Export anything you want to keep.

10Duel, recipes and sharing

Blind duel

Duel mode sends your prompt to two models at once and streams both answers side by side. Names are hidden and sides are shuffled, so neither position nor branding can sway you. Vote A, B or tie, and the models are revealed. Each vote updates a personal ranking (win rate, with ties counted as half) shown on the Duel start screen. Pick Random for both contenders, or pin one model to test it against the field. The answer you vote for becomes the context for follow-ups.

Recipes and the command palette

Press / in an empty message box, or Ctrl/⌘ + K anywhere, to open the palette. Recipes preset the mode, the pair, tools such as web search, and a prompt skeleton:

RecipePairUse for
Code reviewForge → MeridianBugs, security, edge cases, corrected code
Red-team my planAtlas → OrbitWeak assumptions and failure modes
Fact-checkField → Prism, web onClaim-by-claim verdicts with sources
Decision memoAtlas → SummitOptions, criteria, recommendation
Write, then editSummit → MeridianDrafting with an editor pass
Math, verifiedAtlas → FieldSolutions checked by a second method

The palette also switches modes, toggles the challenge stage and web search, exports the chat and jumps to recent conversations. Ctrl/⌘ + Shift + O starts a new chat.

Agreement score

After the challenge stage, the pipeline bar shows how many problems the challenger flagged (✕ fixed) and how many parts of the draft it checked and confirmed (✓ held). Hover it to see the challenger's one-line verdict.

Share links

share under an answer copies a link to a read-only page with your question, the draft, the review and the final answer. The content is compressed into the part of the link after #, which browsers never send to a server, so shared answers are not stored anywhere and cannot be deleted remotely. Anyone opening the link can press Continue in Dyad to import the thread into their own local history. Very long answers are shared without the draft and review to keep the link a usable length.

Dictation

In browsers with speech recognition (Chrome, Edge, Safari), the ● button dictates into the message box in your system language.

11Trace and provenance

Every finished answer carries a trace: job id, timestamp, and for each stage the Dyad name, the engine behind it, input and output tokens, credits and duration. Pair answers also show whether the challenge was skipped and why Auto Pair picked that pair.

The trace shows what ran, not the model's private reasoning; hidden reasoning tokens are excluded from responses.

12HTTP API

The endpoints the app uses are plain HTTP and can be called directly. They are rate-limited per IP and have no auth yet; keyed access with server-side credit balances is on the roadmap.

MethodPathPurpose
GET/api/modelsRoster (names, tiers, vision) and whether live models are connected
POST/api/chatOne streamed completion (SSE, OpenAI-compatible chunks)
POST/api/imageGenerate an image (on the free tier it returns a render URL the browser fetches)
GET / POST/api/storeCredit packs; POST returns a signed 15-minute quote for a wallet
POST/api/purchaseVerify a Solana payment against its quote and return the credits
GET/api/verify-postVerify a public X post for Post & Earn

Streamed completion

curl -N https://trydyad.xyz/api/chat \
  -H 'content-type: application/json' \
  -d '{
    "model": "atlas",
    "temperature": 0.7,
    "length": "medium",
    "messages": [{ "role": "user", "content": "Explain CAP theorem in 5 lines" }]
  }'
# → data: {"choices":[{"delta":{"content":"…"}}]}
# → data: {"choices":[…],"usage":{"prompt_tokens":…,"completion_tokens":…,"cost":…}}
# → data: [DONE]

Body fields: model (a Dyad name), messages, optional temperature, length (short|medium|long), web, pdf. A pair is three calls in sequence, exactly as the app does it — see the stage prompts in section 02.

Errors

{ "error": { "code": "rate_limited", "message": "Too many requests. Try again in a few minutes." } }

Codes: invalid_request, model_unavailable, payload_too_large, rate_limited, provider_error, no_image.

13Post & Earn

Post about Dyad on X, paste the link on /earn and get credits. Verification happens on the server through X's public embed endpoint:

  • the post must be public and exist;
  • its author must match the handle you entered;
  • its text must tag @trydyad or mention $DYAD, Dyad or a link to the site.

Each verified post pays 25 credits, up to 3 posts per day, and each post can be claimed once. Season points (1 per verified post) will feed the leaderboard when the season opens.

14$DYAD

$DYAD is the participation asset: staking, a share of the daily credit pool, and rewards for contributors. You never need it to use Dyad. Full design, allocation and the staking calculator are on the token page.

The token is not live. Contract address, network and final parameters are published only through official Dyad channels at launch. Anyone quoting a contract before that is not us.

15Status and roadmap

SurfaceStatus
Pair modeBeta · live
Chat, attachments, webBeta · live
Image studioBeta · live
Blind duel, recipes, share linksBeta · live
Agents (built-in, custom, shareable)Beta · live
Credit store on Solana (USDC)Live; $DYAD payments at token launch
Post & EarnLive (credits); season leaderboard at launch
Keyed API + server-side balancesNext
$DYAD, staking, credit poolDesign published, contracts at launch
Teams and shared workspacesLater