Anthropic API

Register and Obtain API Key

Create an account on the Anthropic Console and obtain an API key to start integration.

Register at console.anthropic.com, go to API Keys, create a new key and save it to the environment variable `ANTHROPIC_API_KEY`.

Load a small amount into Billing beforehand to avoid disruptions during the first test request.

Make a Simple Test Request

Install the official SDK and send a simple chat request to confirm the connection works.

Run `pip install anthropic`, call `client.messages.create(model="claude-sonnet-4-5", max_tokens=1024, messages=[{"role":"user","content":"..."}])`.

Always set `max_tokens` according to actual needs to avoid unnecessary costs.

Write System Prompt to Define Agent Role

Set up a system prompt clearly describing the role, limitations, and tone of the customer support agent.

Pass the `system` parameter describing: 'You are the technical support assistant for company X, only answering based on the documents provided, always polite and concise.'

Clearly state in the system prompt that the agent must refuse to answer if information is not found, avoiding fabrication of responses.

Connect Document Search Tool via Tool Use

Define an internal document search tool so that Claude can proactively call when additional information is needed.

Pass the `tools` parameter describing the function schema `search_docs(query)`, handle response when `tool_use` is returned by running the actual function and sending back the results.

Limit the number of results returned by the tool (e.g., top 3 document excerpts) to avoid cluttering the context.

Enable Prompt Caching and Monitor Costs

Activate caching for the system prompt and fixed documents to reduce costs for repeated requests.

Mark `cache_control: {"type": "ephemeral"}` for blocks of content that change infrequently in the request, monitor usage on the Console.

Prompt caching is most effective when the cached content is identical across multiple consecutive requests.