Choosing, pricing and delegating with OceanToken

Connect first

These steps use the OceanToken MCP tools (search_models, estimate_cost, generate_image, ...). If they are not available, the server is not connected yet. It is https://mcp.oceantoken.ai/mcp (streamable HTTP) with an OAuth sign-in. In OpenClaw:

openclaw mcp set oceantoken '{"url":"https://mcp.oceantoken.ai/mcp","transport":"streamable-http","auth":"oauth"}'
openclaw mcp login oceantoken

For other hosts see https://github.com/NextFormAI/oceantoken-plugins#install.

Find and compare

• search_models filters by type, publisher, capabilities (vision, tools, reasoning, web_search, audio_output) and max_price, and sort="price" puts the cheapest first. Search short names ("veo", "opus", "qwen") and read ids from the results; never guess an id.
• Prices are in the unit the model is billed in: chat per 1M input / output tokens, images per image (or image tokens), video per second (a "from" price: resolution and audio cost more), speech per 1K characters, transcription per minute.
• When comparing, show a short table: model, price, context window, relevant capabilities, and one line on why each fits.

Estimate before spending

estimate_cost returns USD, a confidence (close, rough, minimum, unknown), the assumptions made and whether it fits the available balance. Pass real numbers when you have them: input_text or input_tokens, output_tokens, seconds + resolution for video, text for speech. Report the estimate with its confidence; do not present a rough or minimum figure as exact.

Delegate with ask_model

ask_model(model, prompt, ...) sends one self-contained request to another model and returns its answer, usage and cost.

Delegating sends the user's text or code to another model provider, so do it only when the user asked for it or agreed to it. Before the first call, say which model you will use and what you will send. Never include secrets, credentials, .env contents or files the user has not chosen to share.

• Good uses, once agreed: bulk or mechanical text (summarise 30 files, translate strings, classify tickets, draft test fixtures) on a cheap model; a second opinion from a different model family on a design or a diff; a long-context model for a document too big to reason over comfortably; a vision model for screenshots (image_urls).
• The delegate sees nothing but your prompt: include the code, the context and the exact output format you want.
• Check what comes back. Treat it as a draft from a colleague, not as truth, and say in your reply that another model contributed.
• Say what it cost when it is not trivial (cost_usd).

Balance

get_account shows balance and available balance (minus holds for running video jobs). If a request is refused for balance, tell the user that credit is added in the OceanToken console, and offer a cheaper model.