What the Claude API is good for
Everything goes through one endpoint, the Messages API. You send instructions, conversation history, documents, images and tool definitions, and Claude replies with text, structured output or tool calls. Tools marked strict always get arguments that match their schema, and structured outputs return JSON that matches the schema you supply.
The features that matter most for business automation:
- Long context. Current models accept up to 1M tokens, enough for full contracts, policy manuals or a large codebase in a single request.
- Prompt caching. Stable context, such as a system prompt, tool list or reference document, is cached and reused across calls, which cuts cost and latency for assistants that send the same material repeatedly.
- Documents and citations. PDFs and images go in directly, and citations tie each statement in an answer to the passage it came from.
- Batches. The Message Batches API processes large jobs asynchronously at a discount.
- Server tools and MCP. Web search, web fetch and code execution run on Anthropic’s side, and the MCP connector lets Claude call remote MCP servers.
- Effort control. Current models think before they answer, and an effort setting trades depth of reasoning against speed and token use per request.
For agents, there are three levels: your own loop over the Messages API, the Claude Agent SDK for a full harness on your infrastructure, and Managed Agents (in beta), where Anthropic runs the loop and a sandbox for you.
How Lumeor uses it
We use the Claude API for steps where reading and judgment matter: contract and document review, drafting replies customers will read, RFP answers and agents that run many tool calls. In n8n we call it through the Anthropic nodes or an HTTP step with a strict schema; in code we use the official SDK with prompt caching on every stable prefix and logging on every call.
We pick the model per step and test it on the client’s real examples. A triage step might run on Claude Haiku 5.5 while the drafting step runs on Claude Sonnet 5.5. See the RFP response guide for a typical design.
Limits to plan around
- Rate limits by tier. Requests and tokens per minute depend on the model and your usage tier. Request increases before a launch, not during it.
- Platform differences. Bedrock, Vertex AI and Foundry carry Claude models but may lag on betas and server tools. Check before committing to a platform.
- Cache discipline. Caching only works if the cached prefix is byte-for-byte identical. Timestamps or reordered tools in the system prompt quietly disable it.
- Long runs take time. Hard agent tasks on the largest models can run for minutes. Use streaming, timeouts and progress updates in your interface.
- Model updates. New models change behavior and some request parameters. Pin model IDs and re-run your evaluation set before upgrading.
When to pick something else
- OpenAI API for real-time speech-to-speech voice and hosted file search. See OpenAI API vs Claude API for agents and the OpenAI API page.
- A framework such as LangChain when you need many providers behind one interface.
- Open-source models when data must never leave your own servers.