The builder layer of Claude, to plug AI into your product or your programmatic workflow.
Who this is for
Devs, tech teams and solo builders who want to build WITH Claude rather than just USE it. If your daily driver is claude.ai or Claude Code, skip this chapter.
If your daily driver is a repo, a backend, a pipeline or a product running in prod, welcome. You're about to expose Claude to your users, or put it to work in the background.
What you can do, 3 distinct tools
Claude API (platform). Direct model calls. You send a prompt, you get a response. The foundation everything else sits on. You get structured outputs (now GA), prompt caching to save on repeated big prompts, streaming for token-by-token output, tools so Claude can call your functions. This is what you use to embed Claude inside YOUR app.
Claude Agent SDK (Python + TypeScript). Reproduces the Claude Code loop in your own environment. Hooks, sub-agents, skills, MCP, the full harness that lets Claude Code plan, run tools, delegate, and verify its own work. If you want to build your own agents with your own scope and your own tools, this is the brick to pick. You keep full control of the infra.
Claude Managed Agents (beta). The version hosted by Anthropic. Execution sandbox, streaming via SSE, infra running 24/7. You write the agent logic, Anthropic runs the machine. For shipping agents without managing servers, retries, or scaling.
How much it costs
Pay as you go, token-based, input and output. Prices depend on the model (Haiku, Sonnet, Opus) and follow the usual shape: the more capable the model, the more a token costs.
Prompt caching genuinely changes the math. On a big recurring system prompt (project instructions, docs, client context), you can push savings up to 90% on the cached chunk. That's the first lever to pull before going to prod.
The Agent SDK and Managed Agents consume the same tokens as the API. Managed Agents add a hosting cost for the managed infra. Exact prices at platform.claude.com.
When to use what
| Tool | You want to... |
|---|---|
| Direct API | Fire a prompt from your app and render the response. |
| Agent SDK | The full loop (planning, tool use, sub-agents, hooks) inside your own code. |
| Managed Agents | An agent running 24/7 without touching infra (watchers, monitors, long-lived jobs). |
Simple rule: as long as one call gets the job done, stay on the API. When you need an agent that thinks, calls tools and comes back with a result, move up to the Agent SDK. And if that agent has to live without you, flip to Managed Agents.
Concrete example: the site gaetanwittebolle.com runs on Resend + Next.js on the product side. A feature like "summarize a chapter on demand" would hit the direct Claude API. A "career agent" watching LinkedIn every day would live on Managed Agents.