Notebooks¶

Three Jupyter notebooks showing the same research question — "Find Argo float profiles around Hawaii in March 2026 where the float reached at least 1500 dbar pressure. Summarize temperature ranges and give me a few download links." — answered by three different LLM providers using their native MCP support.

File	Provider	SDK	API surface
`01-claude-mcp-agent.ipynb`	Anthropic	`anthropic` (Python)	Messages API, `mcp_servers=[...]`
`02-chatgpt-mcp-agent.ipynb`	OpenAI	`openai` (Python)	Responses API, `tools=[{"type":"mcp",...}]`
`03-gemini-mcp-agent.ipynb`	Google	`google-genai` + `mcp` (Python)	`generate_content` with `tools=[session]`

All three notebooks use the Pattern A approach — the LLM autonomously calls AQUAVIEW MCP tools to answer the question. None of them call AQUAVIEW directly from Python; the agent is in the loop.

Why all three?¶

So you can compare. Same MCP server, same prompt, three different agent runtimes — see how each one reasons over the catalog, decomposes the question, and structures its tool calls.

The notebooks intentionally have the same structure: install → setup → first call → inspect tool trace → render answer → multi-turn → streaming → next steps.

Setup (any notebook)¶

pip install jupyter
jupyter notebook

Each notebook has a setup cell at the top that installs the provider's SDK and reads its API key from the environment.

Provider	API key env var	Get a key
Anthropic	`ANTHROPIC_API_KEY`	https://console.anthropic.com
OpenAI	`OPENAI_API_KEY`	https://platform.openai.com
Google	`GOOGLE_API_KEY`	https://aistudio.google.com/apikey

Note on cost¶

Each notebook makes one to three model calls. Token usage is small — the MCP tool calls happen server-side via the AQUAVIEW HTTP endpoint and are billed only for the parts the model actually reads. Expect under $0.05 per full run on each provider.

Differences worth knowing¶

Claude (Anthropic) — mcp_servers is a top-level parameter on messages.create. Tool calls and results appear interleaved in response.content as mcp_tool_use / mcp_tool_result blocks.
ChatGPT (OpenAI) — mcp is a tool type in tools=[...]. Tool calls and results appear in response.output as mcp_call items. Use previous_response_id for multi-turn.
Gemini (Google) — passes the mcp.ClientSession directly into tools=[session]. The SDK auto-discovers tools from the live server and runs the function-calling loop internally.

All three speak the same MCP server underneath. The differences are purely in agent runtime ergonomics.