Claude Cookbooks

What it is

Claude Cookbooks is Anthropic’s repository of Jupyter notebooks and Python examples for building with the Claude API, kept under the MIT license. Its README says the guides offer copy-able code snippets you can drop into your own projects. The examples are mostly Python, and the README says the ideas carry to any language that can call the Claude API. Folders cover tool use, retrieval, classification, summarization, vision, Skills, the Claude Agent SDK, Claude Managed Agents, cost and latency, each with notebooks you can run against your own key.

1. Clone the repository and install its dependencies
2. Copy .env.example to .env and add your API key
3. Pick a recipe folder, such as tool_use or capabilities
4. Open its notebook and run the cells top to bottom
5. Copy the part you need into your own project

Who it suits

Developers who write some Python and want a working example of a Claude API feature before they build their own. The README suggests the Claude API Fundamentals course if you are new to the API. These are teaching recipes. Each one shows a technique, not a finished product.

Skip it if you want a finished app to install, since the Managed Agents README sends its full apps to a separate repository, claude-quickstarts.

Where to use it

The recipes are Jupyter notebooks and Python files, so they run wherever those do, with a Claude API key. The maker’s pages describe these routes:

• Your computer with uv: the repository root, on Python 3.11 or 3.12
• Your computer with pip: the skills folder, in a virtual environment
• A Jupyter kernel from uv: the claude_agent_sdk folder
• Claude Managed Agents, Anthropic’s hosted runtime: the managed_agents notebooks call it with your key
• Claude Code inside a clone: the repository’s review commands and audit skill
• Your browser: read the recipes on GitHub or on the Claude Cookbook site, with no install

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder. It holds a page of steps each for the repository root, the skills, claude_agent_sdk and managed_agents folders, the Claude Code commands and reading online. Start with uv and a clone of the repository.

1. Install uv. Use the shell line or the Homebrew line from CONTRIBUTING.md.
2. Clone the repository and go into it.
3. Install the dependencies. Run uv sync --all-extras.
4. Copy .env.example to .env. Then put your key in it.
5. Open a recipe’s notebook in Jupyter. Run its cells from the top.

git clone https://github.com/anthropics/claude-cookbooks.git
cd claude-cookbooks
uv sync --all-extras
cp .env.example .env

The skills folder’s README uses pip and a virtual environment instead of uv:

git clone https://github.com/anthropics/claude-cookbooks.git
cd claude-cookbooks/skills
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env

How to use it

Pick a folder for what you want to build, open one notebook and run its cells in order. Because the repository keeps saved outputs on purpose, each notebook also shows its results.

CLAUDE.md says a notebook should run top to bottom without errors. The skills README starts Jupyter with this line, after its pip install:

jupyter notebook

The pages give no launch line for the uv route, so open a notebook your usual way. In the Agent SDK folder, this line registers its environment as a kernel:

uv run python -m ipykernel install --user --name="cc-sdk-tutorial" --display-name "Python (cc-sdk-tutorial)"

Example

You want to see how Claude can call many tools without filling its context with their results. You clone the repository, add your key to .env and open tool_use/programmatic_tool_calling_ptc.ipynb. Its page on the Claude Cookbook site is titled “Programmatic tool calling (PTC)” and says the technique lets Claude write code that calls tools in the code execution environment. The notebook uses a mock team expense API. Its query opens with this question, then gives the standard $5,000 quarterly travel budget and asks Claude to check custom budget exceptions for anyone over it:

Which engineering team members exceeded their Q3 travel budget?

It answers first with ordinary tool calls and then with programmatic tool calling, and at the end it prints a table that compares API calls, total tokens and elapsed time. The notebook loads your key from .env. No key sits in the file.

For newcomers and experts

If you are new. The Claude API lets your code send Claude a message and read its reply. A notebook mixes notes with code cells that you run one at a time, and the cells that call Claude need an API key from your Anthropic account. The README suggests Anthropic’s Claude API Fundamentals course first. You can also just read a notebook. Its saved outputs show what each cell printed.

For experts. Recipe 08 in the Agent SDK folder triggers dynamic workflows. Claude writes a JavaScript orchestration script, and a runtime runs it across parallel subagents in the background. Recipe 07 hosts an agent on Docker, Modal and Kubernetes behind one HTTP interface.

• Managed Agents notebooks cover versioned prompts, memory stores, session budgets, Outcomes graders and an advisor roster entry. CONTRIBUTING.md asks each agents.create call to carry an anthropic_cookbook metadata tag
• CLAUDE.md says to use non-dated model aliases. Bedrock IDs differ: use the base Bedrock ID, such as anthropic.claude-opus-5-5, and prepend global. for global endpoints, which CLAUDE.md recommends: global.anthropic.claude-opus-5-5. Integrations that only go through InvokeModel use the dated form
• The Makefile has make test-notebooks for fast structure tests, and make test-notebooks-exec, which runs notebooks and needs an API key
• The audit skill says cookbooks are not production-ready code. They show use cases and capabilities, not production patterns

Good to know

The code and notebooks sit on your machine. Cells that call Claude send your prompts to the Claude API with your key. Files the Skills notebooks generate are stored temporarily on Anthropic’s servers. Managed Agents run your agent in Anthropic’s hosted, sandboxed environment, with the files you mount. Recipe 04 in the Agent SDK folder also calls OpenAI with your OPENAI_API_KEY. CONTRIBUTING.md asks Managed Agents notebooks to tag each agent with an anthropic_cookbook value, so Anthropic can see which cookbooks people run.

• Cost and keys: each cell that calls the API spends tokens on your own Anthropic account, and the pages give no price. Keep your key in .env and never commit it, as CLAUDE.md says. The example env file suggests a cheaper model in CLAUDE_MODEL, for cost savings
• Other services: the third_party folder has recipes for Deepgram, ElevenLabs, LlamaIndex, MongoDB, Pinecone, Voyage AI and Wolfram Alpha. The Managed Agents MongoDB notebook, CMA_with_mongodb_atlas.ipynb, needs your MONGO_URI, and three other Managed Agents notebooks need a GITHUB_TOKEN
• Saved outputs can lag the code: the programmatic tool calling page sets claude-sonnet-5-5 in its code, and its printed output shows claude-sonnet-4-6. Some recipes also use beta headers, so check what a notebook sends
• Dates: the site and registry.yaml give each recipe a date. Check it before you copy older code
• Pages disagree: the kit’s README lists each case. One is the Python version, 3.8 or higher in the skills README and 3.11 or 3.12 in pyproject.toml. Another is a root README link to a file the tree does not hold
• Removal: the pages give no uninstall steps. Deleting the clone removes the code. The Agent SDK route also registers a kernel named cc-sdk-tutorial for your user account, and the pages do not say how to remove it
• License: MIT. The LICENSE file reads “Copyright (c) 2023 Anthropic”