Six starter kits for Claude agents that do real work

What it does Anthropic maintains this monorepo of six quickstart projects, each a concrete starter for a different Claude API use case. The collection covers a customer support agent with knowledge-base access, a financial analyst that chats over interactive charts, a Playwright-driven browser automator, a containerized desktop-control environment, a native macOS computer-use reference implementation, and an autonomous coding agent that checkpoints progress in git. Every project is framed as a foundation to customize rather than a finished product.

The interesting bit The value is in the boring parts. The computer-use best-practices project explicitly demonstrates cost-control and reliability patterns—prompt caching, image pruning, server-side compaction, batched tool calls—that are usually left as an exercise for the reader, and the browser and coding quickstarts provide enough scaffolding to skip the usual reverse-engineering phase.

Key highlights

• Six distinct quickstarts: customer support agent, financial data analyst, containerized computer-use demo, native macOS computer-use best practices, browser automation, and autonomous coding agent.
• The computer-use best-practices project is explicitly pedagogical and demonstrates reliability patterns like explicit tool definitions, correct image sizing and pruning, prompt caching, batched tool calls, and a sandboxed shell.
• The browser-use demo provides a complete Playwright-backed reference implementation for navigating sites, inspecting DOM elements, extracting content, and filling forms.
• The autonomous coding agent implements a two-agent pattern—initializer plus coding agent—that builds applications incrementally across multiple sessions, with progress tracked via git and a feature list.
• All projects are MIT-licensed and require an Anthropic API key.

Caveats

• The top-level README is essentially a directory; the real setup instructions and code details live in each subdirectory, so you cannot evaluate complexity without diving deeper.
• The native macOS computer-use best-practices project is explicitly meant to run against a macOS desktop inside a VM, which may be an immediate barrier for non-Mac developers.

Verdict Worth cloning if you are evaluating Claude for automation workloads and want vendor-sanctioned starting points rather than community snippets. Skip it if you are looking for framework-agnostic abstractions or a single, polished end-to-end product.