How MeowByte works
A short reference for the mechanics behind personalities, the Council, and usage.
Personalities
A personality is the unit you actually talk to. Each one has a name, a system prompt, a model, and generation parameters (temperature, top-p, max tokens). Create as many as you want - a blunt code reviewer and a patient explainer can both exist, backed by different models, without stepping on each other.
Every account starts with a few built-in personalities already configured, so there's something to talk to before you build your own.
Model routing
MeowByte routes every completion through a single model gateway, which means a personality can point at effectively any leading model on the market and be repointed later without losing its system prompt, history, or conversations.
Which models are available to you depends on your tier - see pricing for the breakdown.
The Council
The Council seats two to six personalities around a topic and runs them through five structured phases:
- Restate - each participant independently confirms what's actually being asked, before seeing anyone else's take.
- Analysis - each weighs the idea and its parts on their own, so early opinions don't anchor each other.
- Cross-discussion - participants see each other's analysis and respond, for as many rounds as you configure.
- Synthesis - the group converges on a shared recommendation, with any real disagreement flagged rather than smoothed over.
- Result - one consolidated recommendation, with a shortcut to turn it directly into a new conversation and start acting on it.
The full transcript - every phase, every participant's contribution - stays attached to the council session, and you can pause, resume, or cancel a run in progress.
Usage and tiers
Every message a model generates - chat replies, personality previews while you're editing one, and each council participant's turn - debits the same weekly token quota. The quota resets automatically each Monday at 00:00 UTC; nothing rolls over and nothing expires early.
See pricing for the exact weekly quota and model access on each tier.
API
Call MeowByte directly from your own code, separate from the chat app. API usage is billed from a credit balance you buy in advance ($1 = 100 credits) rather than your weekly chat quota - the two are independent. Generate a key from Settings → API keys.
Want a coding agent rather than raw API calls? MeowByte Code runs on this same API from your terminal.
Each call is charged at the answering model's own price plus a 25% margin,
for input and output tokens separately, so small models cost very little and large ones cost
what they cost. GET /v1/models lists every model's
creditsPerMillionInputTokens and creditsPerMillionOutputTokens, and the
API keys page shows the same table. Charges are exact: a call costing a fraction of a credit is
recorded at its exact cost, and whole credits leave your balance as fractions add up. (Before
October 2026 every model cost a flat 1 credit per 1,000 tokens.)
Base URL: https://backend.meowbyte.net. Every request needs an
Authorization: Bearer <key> header.
GET /v1/models- the current model catalog and pricing.POST /v1/chat/completions-{ model, messages, temperature?, maxTokens?, stream? }. Setstream: truefor a server-sent-events response (delta,model_switched,error, and a finaldoneevent carrying{ usage, balance, model }); omit it for a single JSON response. Every response includes themodelthat actually answered - see below.balanceis your remaining credits after the request, ornullin the rare case it couldn't be read back; the completion is still valid, andGET /v1/creditshas the current figure. A502always means no completion was produced, so it's safe to retry.- Tool calling: pass
tools(OpenAI-style function definitions) and optionallytool_choice("auto","none","required", or a specific function). When the model calls a tool, the JSON response hastoolCallsandfinishReason: "tool_calls"; a stream sends onetool_callsevent with each call complete (arguments already reassembled) andfinishReasonindone. Send the results back as{ role: "tool", tool_call_id, content }messages after the assistant message that carriedtool_calls. These fields use OpenAI's names, so existing tool definitions work unchanged. Only models withsupportsToolsinGET /v1/modelsaccepttools(others return400 model_no_tools). GET /v1/credits- your current balance and recent transactions.GET /v1/usage?days=30- spend broken down by day and by model, for the last 1-90 days (default 30).
A request is rejected with 402 if your balance is at or below zero, or if its worst
case - your input plus maxTokens of output (8,192 if you don't set it) at that model's
price - is more than your balance covers; the response includes estimatedCredits.
Lower maxTokens, send less context, or top up from the API keys page. An unknown model
returns 400 model_not_found, and a model with no published price
400 model_unpriced. Requests are capped at 60 per minute per account, shared by all
of your keys; every response
carries X-RateLimit-Limit/X-RateLimit-Remaining/X-RateLimit-Reset
headers, and going over returns 429 with a Retry-After header.
If the requested model is down or rate-limited, the request automatically retries
once against a reliable default model rather than failing outright - the response's
model field (or the streamed model_switched event) tells you when
that happened, so your code can log or surface it without needing to handle it as an error.
Try it
A real completion, right here, no account needed - a small fixed set of models, a short reply, and a limited number of tries per visitor so this stays free to leave open.
Accounts and usernames
Sign up with an email and password. Your handle follows the same pattern Discord made
familiar - a display name plus a five-digit discriminator, like brad#12345 -
so display names don't have to be globally unique.