# Ocean One API > The single endpoint to the Ocean Platform: models (OpenAI-compatible), tools with > typed invocation contracts, agents, multi-agent workflows, MCP (inbound and > outbound), capability discovery with measured quality scores, and cross-plane > execution traces. Base URL: https://api.plungeai.com — auth is a Bearer key > (`ozk_…` for execution planes, `sk-ocean-…` for model routes). Key facts: the developer guide is now v3.0; the models plane is OpenAI-compatible (use any OpenAI SDK with base_url https://api.plungeai.com/v1); every response carries an x-request-id header that doubles as a trace id; tool execution enforces a trust fence (403 refused / 409 approval_required on money verbs). Chat completions accept `models: [slug,…]` for ordered fallback with retry, outage exclusion, and honest billing of the model that actually served the request — plus guardrails, `@preset/` configs, and an opt-in response cache (see llms-full.txt). ## Docs - [Full developer documentation](https://api.plungeai.com/llms-full.txt): the complete developer guide (v3.0) and route reference as one markdown document — quickstarts, auth, one verified example per plane, error semantics - [OpenAPI 3.1 specification](https://api.plungeai.com/v1/openapi.json): the machine-readable contract for every route - [Human documentation page](https://plungeai.com/docs): the same guide, rendered — also served at https://api.plungeai.com/docs ## AI agents - Runtime integration needs no HTTP client at all: connect an MCP client (Streamable HTTP) to POST https://api.plungeai.com/v1/mcp with a Bearer ozk_ key — the platform self-describes as MCP tools (execute workflows, agents, search the catalog, …) - [MCP developer guide](https://mcp.plungeai.com/docs): every MCP tool, parameter, prompt and resource, client setup, errors and limits — markdown at https://mcp.plungeai.com/docs.md - [PlungeAI agent skills](https://skills.plungeai.com): installable skills that teach an AI coding agent the API, the MCP server and the CLI