Build with AI that stays useful.
The Platform connects a separate OpenAI-compatible API with the focused OSS Studio workspace. They share model access, accounts, safety, and usage.
Quickstart
API keys are optional: anonymous requests work immediately. Create an OSAII Platform API key when you want higher limits, usage attribution, and a stable account identity.
curl -G https://osaii.wyvernhub.net/api/ask --data-urlencode 'q=Explain recursion briefly' --data-urlencode 'model=fast'
Simple ask
A lightweight GET endpoint for scripts, links, cron jobs, and quick integrations. Use text, json, ccjson, or rjson with the format parameter.
q is required. model defaults to fast and accepts smart, mini, or an exact model ID.
Ask about docs
Ask a question about this documentation. This works without a key; add one when you want authenticated limits and usage attribution.
Chat Completions
Use the OpenAI-compatible endpoint for streaming, tools, reasoning controls, structured outputs, and multimodal messages where supported.
Set stream:true for server-sent events. Tool definitions, tool results, reasoning fields, and structured-output fields are forwarded when supported.
curl https://osaii.wyvernhub.net/api/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"poolside/laguna-xs-2.1","messages":[{"role":"user","content":"Hello"}]}'Credits & pricing
All API models currently cost $0 for input, output, cached input, and requests. 1 credit = $1 USD. Credit purchases are disabled, and a zero balance does not block API access. API keys are optional; existing rate limits and safety rules still apply.
Use the API dashboard to manage named keys, view account API usage, and check your credit balance. Anonymous usage is not attached to an account. Token counts depend on provider reporting.
GET /platform/api/pricing publishes the free pricing policy. Signed-in Platform sessions can read GET /platform/api/dashboard?days=30 (7 or 30 days, UTC) and GET /platform/api/credits. Credit mutations are disabled server-side.
Cozy Swarm
Cozy serves locally ingested open-weight models through a native coordinator and voluntary workers. OSAII exposes its deployment boundary at /cozy/api: scoped Platform keys, account-bound enrollment, mandatory input and final-output screening, and signed coordinator requests. The Worker does not execute model weights.
A Platform key needs cozy:read; generation requires cozy:inference. Create a scoped key through POST /cozy/api/keys while signed in. Existing Platform keys have no Cozy access until explicitly scoped.
Workers use authenticated outbound WebSockets, verified safetensors shards, coarse consecutive layer partitions, explicit TRUSTED or COMMUNITY privacy classes, and bounded failover. Start at the Cozy Swarm page.
Responses
Use the OpenAI Responses-shaped endpoint when your client sends input and expects output or Responses streaming events.
Set stream:true for server-sent events. The gateway preserves the request and response shape supported by the selected model. Text and input_image inputs are accepted on /api/v1/responses; ZeroTwo vision models serve Responses through the chat adapter, while Poolside models remain text-only. When configured, osaii/demo-ollama-random routes through this API to the mini PC, which selects an installed local Ollama model per request. The stream flag is preserved, and x-osaii-demo-local-model identifies the selected backend model.
curl https://osaii.wyvernhub.net/api/v1/responses -H "Content-Type: application/json" -d '{"model":"poolside/laguna-s-2.1","input":"Hello"}'Models & limits
Use catalog IDs with the OpenAI-compatible API. Studio also provides an Auto option that routes between Fast and Smart.
microsoft/bitnet-b1.58-2B-4T
CPU streaming by default. Set fast_mode:true for the demo GPU path; GPU streaming is unavailable because the upstream demo returns only completed jobs.
poolside/laguna-xs-2.1
Low latency.
poolside/laguna-s-2.1
More capable reasoning.
Kimi K3, MiniMax M3, DeepSeek V4 Flash, and DeepSeek V4 Pro.
Studio task-based routing.
Safety boundary
Safety scanning is currently disabled for all model families and endpoints.
General safety filtering before inference.
Final responses checked before delivery.
Advanced requests also pass the under-18 screen.
Studio opens a dedicated support experience.
Studio: the hands-on side of the Platform
Persistent memory, searchable chats, project knowledge, attachments, research, Canvas, data analysis, assistants, scheduled tasks, browser tools, voice, and managed agent sessions.
MythOS desktop integration
MythOS is a separate desktop operating-system product. It works without an OSAII account or OSAII applications.
The optional osaii-web-apps System Hub feature adds browser shortcuts for Platform, Studio, and Masque. They are absent by default and removable.
Local users, files, administrator credentials, and diagnostics never become Platform account data. Browser sign-in is explicit.
GET /mythos/release.json publishes version, channel, architecture, base, and canonical service routes.
Build validation does not imply broad hardware acceptance. The preview currently targets amd64 PC and Intel Mac boot paths.
Masque character platform
Masque is a separate 18+ character product that shares OSAII Platform accounts, model infrastructure, safety, usage, and API keys with Studio and the API.
GET /masque/api/v1/characters?q=detective
POST /masque/api/v1/sessions
POST /masque/api/v1/sessions/{session_id}/messagesMasque exposes Auto, Fast, and Smart—not the general OpenAI-compatible model catalog. Public API keys use explicit masque:read, masque:chat, masque:write, masque:embed, and masque:webhook scopes. Existing keys begin with no Masque privileges.
Character resources include revisions, collaborators, remixes, imports, exports, personas, memory, rooms, social actions, reports, embeds, notifications, and aggregate webhooks. Message responses use SSE events named start, delta, finish, and error, with session, message, branch, revision, selected family, finish reason, and usage fields.
Cosaii local agent
Cosaii (pronounced cozy) is the Platform's local-first developer agent: a downstream fork of Nous Research's Hermes Agent with OSAII device login, scoped Platform inference, strict approvals, and the Agent Skill Repository (ASR) capability system. The agent runtime lives at github.com/osai-initiative/cosaii; this service provides its cloud boundary.
curl -fsSL https://osaii.wyvernhub.net/install/cosaii.sh | sh
PowerShell: irm https://osaii.wyvernhub.net/install/cosaii.ps1 | iex
Run cosaii login. The CLI opens the normal Platform authorization page and polls a ten-minute device grant; no localhost callback is opened.
Conversations, sessions, and learned ASR capability pins stay on the local machine and are not synchronized into Studio transcripts.
Deny All, Allow Low-Risk, Always Ask, and Full Access. Risky actions require explicit approval cards unless the user deliberately relaxes the mode.
cosaii # open the TUI cosaii login # device authorization cosaii ask "fix the failing tests" cosaii asr search "rust async"
Cloud endpoints are scoped under /platform/api/cli/* and /cosaii/api/v1/*. These require CLI bearer tokens and are not additions to the public OpenAI model list. The relay transports opaque ciphertext only.
Lake repositories
Lake is the separate OSAII repository hub for Models, Datasets, Papers, and static/WASM Apps. It shares Platform accounts but has its own Git authorization, R2 object storage, moderation, audit trail, and feature flags.
GET /lake/api/v1/repositories
POST /lake/api/v1/repositories
GET /lake/api/v1/repositories/{namespace}/{slug}/releases/{tag}Public repositories are searchable and anonymously readable. Unlisted repositories work at their canonical URL but are excluded from discovery. Private repositories require Lake membership. Git and LFS use scoped Platform keys: lake:read, lake:write, and lake:admin. Existing keys have no Lake privileges until explicitly granted.
Apps are static/WASM only; Lake never runs uploaded code. Browser runners and viewer-supplied endpoints are opt-in and execute or store credentials locally in the viewer's browser.