Claude Agent SDK App

All three run an agent on Claude; they differ in how much comes built in and who runs it. The tool runner in Anthropic's client SDK is the thinnest: it loops over tools you define and nothing else, with no file or shell tools and no sessions, and it runs in any process, including serverless functions. The Agent SDK, this stack, adds Claude Code's built-in tools, subagents, hooks, permissions, and sessions, at the price of running a subprocess in a container you operate. Managed Agents moves the loop and the sandbox to Anthropic, configured through the API, so there is no container to run, but the agent runs on Anthropic's platform under its controls. Pick the SDK when the agent needs a real shell and files and you want to own where it runs.

It follows the work. One-off tasks, such as fixing a bug or extracting an invoice, suit a fresh sandbox that is destroyed afterwards, which gives clean isolation and no state to clean up. An agent serving continuous traffic, such as a chat bot or an email triager, suits a long-running container holding many sessions warm. Sessions that pause for hours, such as research that resumes later, suit the hybrid: a container that shuts down when idle and reloads the transcript from a session store when the user returns, which makes the store mandatory. The agent runs on one of the two kinds of platform, a container host or a sandbox platform, and they map to these patterns: Railway and Hetzner for long-running pools, Modal and E2B for per-task sandboxes, Fly.io and Daytona for the hybrid.

Tokens, not hosting. Every turn resends the instructions, tool definitions, and conversation so far, and file and shell tools pull long outputs into the context, so a long session costs more with each step. Cap turns per session, give subagents a cheaper model for simple steps, keep tool output short, and track cost per session with the SDK's accounting; the Pricing section has the per-token rates. A Claude Pro or Max subscription can't pay for a product built on the SDK: Anthropic's docs say products may not offer claude.ai login or its rate limits unless approved, so billing goes through an API key, or through Bedrock, Google Cloud, or Foundry on your cloud account.

TypeScript AI Agent builds the agent with a framework, the Vercel AI SDK by default, inside a Next.js app that also renders the chat, and it can switch between Claude, OpenAI, and Gemini. It runs well on serverless hosting, and you write each tool yourself. This stack starts from Claude Code's harness instead: file editing, shell commands, web search, subagents, and sessions come built in, but it is Claude only and each session needs a container. Pick TypeScript AI Agent for a chat assistant inside a web product, and this stack for an agent that works on files and runs commands, or when you want a harness proven on coding work rather than one you assemble.

Yes, and the agent service stays separate from it. The container exposes an HTTP or WebSocket endpoint, keeps each user's session mapped to its subprocess, and streams messages back as the agent works. A Next.js app, such as the interface in the TypeScript AI Agent stack, or any other web frontend calls that endpoint and renders the stream, including tool calls. Anthropic's guidance puts user authentication at a gateway in front of the agent container, so the frontend and the gateway handle sign-in and the agent only sees requests that are already authorised. For agents with no interface, a queue or a scheduler can call the same endpoint.

Yes. The Python package has the same loop, tools, hooks, subagents, MCP support, and session stores, needs Python 3.10 or later, and bundles the same binary; it is also the MIT-licensed one of the two. The differences are in the API surface: Python holds a session open with a client object where TypeScript streams input into a running query, and a few helpers, such as pre-warming subprocesses, exist only in TypeScript. Everything else in this stack carries over: the hosts, the session stores, and the tracing. Choose the language your team and your tools already use.