A project by Anthropic (
anthropics/claude-quickstarts, MIT).
Why it matters
Building an application on the Claude API means writing more than API calls: a UI, tool definitions, session handling, and integrations with other services. Claude Quickstarts gives you runnable starting points so you don’t build those apps from scratch. In the README’s words, each quickstart provides “a foundation that you can easily build upon and customize for your specific needs.”
What you can build
The 18 quickstarts in the root README, grouped by topic:
Claude API apps
- Customer Support Agent — an AI-assisted customer support system with access to a knowledge base
- Financial Data Analyst — analyze financial data via chat with interactive data visualization
Computer use and browser automation
- Computer Use Demo — an environment and tools that Claude can use to control a desktop computer, running in a Docker container
- Computer Use Best Practices — a native-macOS reference implementation (run it in a VM) showing explicit tool definitions, image sizing and pruning, prompt caching, server-side compaction, batched tool calls, a sandboxed shell, and trajectory recording
- Browser Use Demo — a reference implementation for browser automation that navigates websites, interacts with DOM elements, extracts content, and fills forms using a Playwright-backed browser tool
- Browser Toolset, Computer Toolset — minimal Python and TypeScript examples of
browser_toolset_20260801andcomputer_toolset_20260801. The README says they show the interface and are not production code
Claude Agent SDK
- Autonomous Coding Agent — a two-agent pattern (initializer + coding agent) that builds complete applications over multiple sessions, with progress persisted via git and a feature list
Claude Managed Agents
- Chat SDK, CopilotKit with AG-UI, Road Trip Planner — browser chat apps that stream replies token by token, show tool calls in a live feed, or render them as interactive generative UI
- Slack, Linear — stateless webhook bridges that answer
@mentionsin a thread or as a comment - Daily Brief — a scheduled agent that reads your Slack channels and GitHub pull requests each weekday and posts one short brief to Slack
- Knowledge Wiki — distills a document corpus into a versioned memory-store knowledge wiki and answers repeated analyst questions with provenance on every fact, using public SEC EDGAR filings as the worked example
- Sentry, Sentry Autofix — scheduled Sentry triage, or a webhook-driven session per issue that reproduces the bug with a failing test, pushes a branch, and opens a pull request for a person to review
- MCP Server (TypeScript) — exposes the Managed Agents Sessions API as nine tools so Claude Desktop, Claude Code, or claude.ai can drive the agents already in your workspace
The managed-agents/ folder also has assistant-ui (a spreadsheet analyst in a browser chat) and self-hosted-sandboxes (sessions on hardware you control), which are not listed in the root README. For Managed Agents itself, see Claude Managed Agents.
Quickstarts that use Claude Code
Most quickstarts are standalone apps. These ones use Claude Code directly:
- Autonomous Coding Agent — requires the latest Claude Code CLI and the Claude Agent SDK.
- Managed Agents MCP Server — after
bun install, runclaude "walk me through setting this up"and Claude readsskill.mdand drives the rest. To add it by hand, register the stdio server withclaude mcp add managed-agents -- bun .... - Sentry, Sentry Autofix —
start.shinstalls Sentry’s Agent Plugin for Claude Code at project scope, then Claude Code walks you through picking the org and project.
How to use
The README’s General Usage steps:
- Clone this repository.
- Navigate to the specific quickstart directory.
- Install the required dependencies.
- Set up your Claude API key as an environment variable.
- Run the quickstart application.
git clone https://github.com/anthropics/claude-quickstarts.git
cd claude-quickstarts/financial-data-analyst
npm install
npm run dev
Requirements differ per quickstart: Node.js 20+ for Customer Support Agent and Financial Data Analyst, Docker for Computer Use Demo, Python and the Claude Code CLI for Autonomous Coding Agent, and Bun with the ant CLI 2.0 or later for the Managed Agents quickstarts. Check each folder’s README for the exact commands.
Notes
- Cost — The code is MIT licensed, but running it uses the Claude API and is billed. Sentry Autofix gives each session a hard spend cap (
AUTOFIX_MAX_USD, default $10 at list prices). - Computer use risks — The Computer Use Demo README warns that computer use poses unique risks distinct from standard API features. It recommends a dedicated virtual machine or container with minimal privileges, keeping sensitive data such as account login information away from the model, and asking a human to confirm decisions with meaningful real-world consequences. It also notes that instructions on webpages or in images may override user instructions (prompt injection).
- Run time — The Autonomous Coding Agent targets 200 features by default, and the README warns that building all of them typically requires many hours of total runtime across multiple sessions. Reduce the feature count for a quicker demo.
- MCP server exposure — If you deploy the Managed Agents MCP server over HTTP, its bearer token lets the holder list, read, and run every agent in the workspace. The README says to set
ALLOWED_AGENT_IDSandALLOWED_HOSTSbefore you deploy. - Minimal examples — Browser Toolset and Computer Toolset show the interface and are not production code.
- Further resources — The README points to Claude Cookbooks (code snippets and guides for common tasks) and the Claude API Fundamentals Course.