Jonathon's AI Wiki

anthropic/claude-quickstarts — Official Starter Apps for the Claude API

4 min read

Source: gh-star-anthropics-claude-quickstarts.md

Repo: https://github.com/anthropics/claude-quickstarts Stars: ~16,715 (at time of ingest 2026-05-20) Language: Python License: MIT Last pushed: 2026-05-13 Maintained: Anthropic (official)

Official Anthropic-maintained collection of ready-to-deploy starter applications built on the Claude API. Each project is a self-contained reference architecture demonstrating a common production pattern — customer-support assistant, financial-data analyst, computer-use demo, agent quickstart — that a developer can fork, swap in their API key, and ship as the skeleton for a real application. Sits alongside the Anthropic Cookbook as the second canonical Anthropic-published reference surface — cookbook is notebook-style demos of a single technique, claude-quickstarts is full deployable projects.

Key Takeaways

  • Official, MIT, actively maintained. Pushed 2026-05-13, ★16.7k. Under the anthropics/ org alongside anthropics/skills, anthropics/anthropic-cookbook, and anthropics/anthropic-sdk-python — the de-facto first-read pile for any new Claude API project.
  • Full-application starters, not snippets. Each subdirectory is a complete project with requirements.txt / package.json, README, environment-variable scaffolding, and ready-to-run entry point. Distinct from the cookbook’s notebook-per-technique format.
  • Headline projects (as of ingest):
    • Customer support agent — retrieval-augmented conversation loop with tool calling for ticket lookup / order status / FAQ search. Reference architecture for the most common production Claude use case.
    • Financial data analyst — pandas + matplotlib + Claude for natural-language analysis over uploaded CSV/XLSX. Pairs with Computer Use when extended to ops dashboards.
    • Computer use demo — the reference container + sandbox setup for Computer Use before it landed as a Claude Code primitive. Still the canonical environment for testing Computer Use against arbitrary applications.
    • Agent quickstart — minimal tool-use loop with a small custom-tool catalog. Best first read for the basic Messages-API-plus-tools shape that the Claude API exposes underneath all agent workflow patterns.
  • Use it as a teaching artifact. The repo is the answer to “what does an Anthropic-blessed production Claude API loop look like?” The customer-support and agent-quickstart projects in particular show the boundary between what the SDK does and what application code needs to handle (rate-limit retries, tool result formatting, system-prompt structure).
  • Why this matters for WEO Marketly. When prototyping a new Claude API project (Hermes tool, internal SaaS feature, custom MCP server), fork the closest quickstart first instead of writing the boilerplate. Saves 1-2 hours of glue code and inherits Anthropic’s current best-practice retry/error-handling shape.

How it composes with the Claude API stack

  • vs Anthropic Cookbook: Cookbook = single-technique notebooks (prompt caching, JSON mode, multimodal). Quickstarts = full applications composing many techniques. Read the cookbook first to learn a primitive, then steal from quickstarts to ship something using it.
  • vs skills: Skills = portable agent capabilities that load into Claude Code / Cowork / Managed Agents. Quickstarts = Messages-API applications that call Claude directly. Different surfaces.
  • vs Build Your Own Coding Agent: The third-party tutorial walks a from-scratch coding agent. Quickstarts’ agent-quickstart is the Anthropic-blessed minimum viable version of the same loop.

Implementation

  • Tool/Service: anthropics/claude-quickstarts (GitHub, MIT)
  • Setup:
    1. git clone https://github.com/anthropics/claude-quickstarts
    2. cd into a project subdirectory (e.g. customer-support-agent/)
    3. Set ANTHROPIC_API_KEY in your shell or a .env
    4. pip install -r requirements.txt (or npm install) per the project README
    5. Run the entry script (python main.py / npm start)
  • Cost: Free repo. Running the projects burns Anthropic API tokens (the API key, not the Claude Code subscription — they are separate billing surfaces).
  • Integration notes: Several projects assume specific model IDs that may drift. Before forking, bump model strings to the current generation (Opus 4.7 / Sonnet 4.7 / Haiku 4.5) and re-read the prompt caching configuration — the repo’s caching examples sometimes lag a release cycle behind the SDK’s current defaults.

Try It

  • Customer support prototype — fork customer-support-agent/, replace the toy ticket database with a connector to GHL / HubSpot / Intercom (see GoHighLevel), keep the conversation loop and tool definitions.
  • Financial analyst for internal data — fork financial-data-analyst/, point it at the WEO Marketly Stripe export or GA4 export, swap matplotlib for Plotly if the output needs to be embedded in a Cowork Live Artifact dashboard.
  • Agent-loop teaching artifact — read agent-quickstart/ before writing any new Claude API tool-use loop in another project. The structure encodes the current Anthropic-blessed retry + tool-result-formatting shape.

Open Questions

  • Does the customer-support agent quickstart use the current prompt-cache shape (ephemeral cache control) or the older inline caching? Verify on next fork.
  • Are any of these quickstarts about to be deprecated in favor of Managed Agents reference apps? The cookbook has been the leading edge of Managed Agents reference material; quickstarts may follow.

Graph View

Backlinks