Features
Everything on this page works today. What each plan includes is on the pricing page.
AI router
One endpoint and one key in front of the providers you already pay for, so your coding tools and apps keep working when one of them fails.
One endpoint for OpenAI and Anthropic clients
Point a tool at the gateway and it speaks Chat Completions on /v1/chat/completions or Anthropic Messages on /v1/messages, with /v1/models to list what a key can use. When the request already matches the provider’s format it is forwarded as sent, with only the model replaced. When it does not, text, tools and tool results are translated, and content that cannot be carried is rejected with an error that names the field.
Your own provider keys
Connect Anthropic, OpenAI, Google Gemini, Groq, OpenRouter, DeepSeek, Mistral and Cerebras, an Ollama server the cloud can reach, or any OpenAI-compatible endpoint with your own base URL. Keys are stored encrypted and only their last four characters are ever shown. Your providers bill you for what you use with your own keys. On Pro and Team, cloud assistant runs can also use the model credit that comes with the plan, where it is switched on.
Router keys with their own limits
Create keys that start with sk-bky-. Only a hash is stored and a key is shown once, when you create or rotate it. Each key can have its own requests per minute, monthly token cap, monthly dollar budget, allowed models and expiry, and can be revoked at any time.
Fallback combos
Save an ordered list of up to 8 connection and model steps and choose how they are tried: in priority order, round robin, cheapest first, or by remaining headroom. A provider that keeps failing is set aside for a while, a rate-limited one waits out its retry time, and the router moves to the next step. It never switches provider in the middle of an answer. Use the model name auto to let the router choose.
Usage you can read
See requests, tokens, failures, fallbacks, latency and cost at list price, grouped by key, model or provider. Request records hold metadata only, never prompt or reply text, and are deleted after 90 days.
Bokyaa as an MCP server
Point an MCP client at https://api.bokyaa.com/v1/mcp with a router key that has the mcp scope. The tools are list_models, get_usage, list_assistants, ask_assistant, search_memory and list_pending_approvals, and add_memory when the key also has the memory:write scope. There are no web search or fetch tools, and nothing here can decide an approval or return a key. Each call is recorded with its tool name and outcome, never its text.
A router on your Mac
The next Mac build includes a local router, off until you turn it on. It listens on the Mac only.
Assistant
The voice-first assistant for you and your business, with every action shown before it happens.
Voice-first Bokyaa
Talk to it like a person. Hands-free or push-to-talk, spoken replies, and you can cut in mid-sentence with headphones or external speakers. On Pro and Team, cloud assistants remember facts you approve across conversations.
Works on your Mac, with your approval
Opens apps, reads and writes files in folders you allow, manages windows and runs your Shortcuts. Anything that changes something asks first by default; destructive actions always ask, and every action is logged.
Private by default
On the Community plan the Mac app runs on-device models and the voice loop on your Mac, so that audio and those prompts never leave it, unless you add your own cloud key to the local router: then the requests it sends to that provider go there. Cloud dictation and server-generated voices are not switched on yet: dictation is unavailable, and when the assistant has voice turned on, spoken replies in the web app and the website widget are read aloud by your own browser’s voice on your device, and only by a voice the browser reports as local. Paid plans add cloud runs when you choose them. Error and crash reports from your apps are off by default; the privacy page lists what they contain.
Your own branded assistant
Give an assistant its own name, logo, colours, persona and allowed topics, then offer it as a text chat widget on your website and to your staff in the app.
Tools through MCP
Connect Model Context Protocol servers so cloud assistants can use your tools. The cloud connects to servers over the network (Streamable HTTP) and never runs a server as a local program.