[{"data":1,"prerenderedAt":302},["ShallowReactive",2],{"en-post-154":3},{"id":4,"title":5,"body":6,"description":294,"extension":295,"meta":296,"navigation":297,"path":298,"seo":299,"stem":300,"__hash__":301},"content/post/154.md","154",{"type":7,"value":8,"toc":282},"minimark",[9,18,23,36,40,88,102,106,137,140,144,220,223,227,230,234,257,260,264,267,271,274],[10,11,12,13,17],"p",{},"Building a real app on the Claude API rarely starts with a prompt. It starts with arguing about folder layout, where session state lives, and how tool calls should be wired. The docs explain each capability clearly, but assembling them into one working project still means trial and error. ",[14,15,16],"code",{},"claude-quickstarts"," turns Anthropic's own recommended patterns into complete projects you can clone and deploy after changing a config file. It does not teach concepts. It hands you code that already runs.",[19,20,22],"h2",{"id":21},"what-is-claude-quickstarts","What Is claude-quickstarts",[10,24,25,27,28,31,32,35],{},[14,26,16],{}," is Anthropic's official collection of starter projects, built to get a deployable Claude API application running with as little boilerplate as possible. The repository was created on 2024-08-29, still had commits on 2026-11-06, and has roughly 20,104 stars and 3,320 forks under the MIT license. Every quickstart is a standalone project rather than a snippet, covering customer support, financial data analysis, computer use, browser automation, and autonomous coding. An ",[14,29,30],{},"agents/"," directory groups them, and a shared ",[14,33,34],{},".pre-commit-config.yaml",", ruff setup, and test gates keep the engineering consistent. That is more structure than most example repositories bother with. For teams moving official capabilities straight into a product, it removes the week spent inventing a scaffold. That structure also makes it easy to diff two examples.",[19,37,39],{"id":38},"features","Features",[41,42,43,47,54,61,64,67,70,85],"ul",{},[44,45,46],"li",{},"Support and analysis. Customer Support Agent ships with a knowledge base, and Financial Data Analyst produces charted analysis. Both are full projects you can edit.",[44,48,49,50,53],{},"Computer use. Computer Use Demo and Best Practices show a desktop agent driven by ",[14,51,52],{},"computer_toolset_20260801"," over VNC and RFB.",[44,55,56,57,60],{},"Browser automation. Browser Use Demo and Browser Toolset use ",[14,58,59],{},"browser_toolset_20260801"," on top of Chrome DevTools Protocol.",[44,62,63],{},"Autonomous coding. Autonomous Coding Agent splits the work into an initializer and a coding agent.",[44,65,66],{},"Managed Agents examples. Chat SDK, Daily Brief, Knowledge Wiki, Linear, Slack, Sentry, Sentry Autofix, Road Trip Planner, and an MCP Server all live here.",[44,68,69],{},"Modern frontend references. Vercel Chat SDK, the AG-UI protocol from CopilotKit, and Next.js show up in the interfaces.",[44,71,72,73,76,77,80,81,84],{},"Pinned dependencies. ",[14,74,75],{},"anthropic 2.0",", ",[14,78,79],{},"@anthropic-ai/sdk 1.0",", and ",[14,82,83],{},"ant 2.0"," keep examples reproducible.",[44,86,87],{},"Engineering discipline. Pre commit hooks, ruff, and test gates come with the repository.",[10,89,90,91,96,97,101],{},"For the browser piece on its own, see ",[92,93,95],"a",{"href":94},"/post/140/","What Is agent-browser and How to Use It",". For skill authoring, ",[92,98,100],{"href":99},"/post/121/","What Is anthropics/skills and How to Use It"," is the reference.",[19,103,105],{"id":104},"how-to-use","How to Use",[107,108,109,112,118,121,128,131,134],"ol",{},[44,110,111],{},"Open the repository, clone it locally, or fork it to your own account.",[44,113,114,115,117],{},"Pick one scenario under ",[14,116,30],{}," and read its README before running the whole repository.",[44,119,120],{},"Set up the stack, a Python venv or Node dependencies for TypeScript.",[44,122,123,124,127],{},"Copy the example environment file, add credentials such as ",[14,125,126],{},"ANTHROPIC_API_KEY",", and never commit the key.",[44,129,130],{},"Start the service per the README and run the default example to confirm toolset and model versions agree.",[44,132,133],{},"Change prompts, tool definitions, and data sources, then check the result with the bundled tests and ruff.",[44,135,136],{},"Follow the project's deployment notes to push to a host. Most examples already include build configuration.",[10,138,139],{},"Toolset names carry a date suffix, so check the pinned version before upgrading. Mixing versions tends to fail on argument validation.",[19,141,143],{"id":142},"supported-platforms","Supported Platforms",[145,146,147,160],"table",{},[148,149,150],"thead",{},[151,152,153,157],"tr",{},[154,155,156],"th",{},"Item",[154,158,159],{},"Detail",[161,162,163,172,185,193,204,212],"tbody",{},[151,164,165,169],{},[166,167,168],"td",{},"Languages",[166,170,171],{},"Python and TypeScript",[151,173,174,177],{},[166,175,176],{},"Core dependencies",[166,178,179,76,181,76,183],{},[14,180,75],{},[14,182,79],{},[14,184,83],{},[151,186,187,190],{},[166,188,189],{},"Capabilities",[166,191,192],{},"Claude Agent SDK and Claude Managed Agents",[151,194,195,198],{},[166,196,197],{},"Toolsets",[166,199,200,76,202],{},[14,201,52],{},[14,203,59],{},[151,205,206,209],{},[166,207,208],{},"Frontend",[166,210,211],{},"Vercel Chat SDK, AG-UI, Next.js",[151,213,214,217],{},[166,215,216],{},"License",[166,218,219],{},"MIT",[10,221,222],{},"Computer use examples need a VNC environment and browser examples need Chrome with CDP, so install those first. Windows, macOS, and Linux all work, with Node and Python versions set by whichever example you choose. A local run needs only the runtime for that example.",[19,224,226],{"id":225},"pricing","Pricing",[10,228,229],{},"The repository is MIT licensed, so cloning, modifying, and shipping it commercially cost nothing and need no account. The real expense is calling the Claude API, billed per token by Anthropic, whether you use these examples or not. Set a usage cap before local testing. Browser and computer use examples loop through multiple calls on their own, and a runaway loop burns budget quickly. Third party services used by an example bill separately under their own terms.",[19,231,233],{"id":232},"use-cases","Use Cases",[41,235,236,239,242,245,248,251,254],{},[44,237,238],{},"First Claude API project. Walk one complete implementation instead of assembling documentation fragments.",[44,240,241],{},"Team scaffolding. Adopt the official structure as a baseline for directories, dependencies, and tests.",[44,243,244],{},"Support or knowledge base prototypes. Customer Support Agent already connects retrieval to answers.",[44,246,247],{},"Trying computer and browser use. See how the official code calls the toolsets without building VNC or CDP plumbing yourself.",[44,249,250],{},"Studying agent orchestration. Two agent coding and MCP Server patterns both have runnable implementations.",[44,252,253],{},"Evaluating new toolsets. The 2026-08 computer and browser toolsets are best understood through these examples.",[44,255,256],{},"Comparing releases. New toolset versions arrive as separate examples, so the changes stay visible.",[10,258,259],{},"Treat these as starting points, not production frameworks. Permission controls, logging, and cost limits still need to be added for your own context.",[19,261,263],{"id":262},"how-it-works","How It Works",[10,265,266],{},"Anthropic's own team maintains the repository as an authoritative reference rather than a product framework, so updates follow capability releases closely. When a new toolset lands, examples tend to appear soon after. The projects sit side by side and can be forked independently, which means one broken demo does not affect the rest. MIT licensing lets downstream teams turn an example into a commercial product, while Anthropic gets a cheaper onboarding path for developers. That is a familiar developer relations trade, and it works because the code is useful on its own. The valuable part is not the line count. It is the undocumented engineering decisions, such as where session state lives and how a failed tool call should retry.",[19,268,270],{"id":269},"conclusion","Conclusion",[10,272,273],{},"If you have read the API docs and still do not know what your first project should look like, start with the example closest to your scenario. Each one carries complete dependencies and tests, so running the default flow before adding your own logic is steadier than stitching snippets together. Check which SDK version an example pins before you begin, and avoid mixing projects from different release periods.",[10,275,276,277],{},"Project ",[92,278,279],{"href":279,"rel":280},"https://github.com/anthropics/claude-quickstarts",[281],"nofollow",{"title":283,"searchDepth":284,"depth":284,"links":285},"",2,[286,287,288,289,290,291,292,293],{"id":21,"depth":284,"text":22},{"id":38,"depth":284,"text":39},{"id":104,"depth":284,"text":105},{"id":142,"depth":284,"text":143},{"id":225,"depth":284,"text":226},{"id":232,"depth":284,"text":233},{"id":262,"depth":284,"text":263},{"id":269,"depth":284,"text":270},"Building a real app on the Claude API rarely starts with a prompt. It starts with arguing about folder layout, where session state lives, and how tool calls should be wired. The docs explain each capability clearly, but assembling them into one working project still means trial and error. claude-quickstarts turns Anthropic's own recommended patterns into complete projects you can clone and deploy after changing a config file. It does not teach concepts. It hands you code that already runs.","md",{},true,"/post/154",{"description":294},"post/154","bGSmCSixKh7ryL9cXm9wT5hUpTKWCgpq9eEJlBq4tIE",1794378304903]