Botmonster TechAnthropic's cookbook hides 91 notebooks behind a list
The Claude Cookbooks are Anthropic’s official notebooks for building with Claude. The repo holds 91 of them, while the README’s table of recipes points at maybe a dozen. You won’t find the 15 managed-agent notebooks, the 8 agent SDK notebooks, or the whole skills folder in that table.
Key Takeaways
- The repo holds 91 runnable notebooks, far more than the README’s dozen.
- The newest material covers managed agents, the agent SDK, and skills.
- The table of contents is stale, so browse the folders instead.
- Two notebooks still call retired models and will fail on a fresh run.
- You need your own API key, and running the notebooks costs tokens.
What is in the Claude Cookbooks today
The Claude Cookbooks repository carries 91 Jupyter notebooks across about a dozen top-level folders. The prompt recipes it started with are now a minority, outnumbered by agent engineering.
| Directory | Notebooks | What you get |
|---|---|---|
| managed_agents | 15 | Agents that run on Anthropic’s hosted sandbox |
| misc | 14 | Single-topic recipes: caching, evals, JSON mode, citations |
| tool_use | 14 | Tool loops, parallel calls, context compaction |
| third_party | 13 | Pinecone, MongoDB, LlamaIndex, Deepgram, ElevenLabs, Wikipedia |
| claude_agent_sdk | 8 | Building and hosting agents with the official SDK |
| capabilities | 6 | Classification, retrieval, summarising, text to SQL |
| multimodal | 6 | Vision, charts, forms, sub-agents |
| patterns | 4 | Workflow shapes: orchestrator, evaluator, async fan-out |
| skills | 3 | Skill packages for Excel, PowerPoint and PDF work |
| extended_thinking | 2 | Thinking mode, with and without tools |
| coding, evals, observability, tool_evaluation, finetuning, fable_5_fallback_billing | 1 each | Newer one-off ideas |
The managed agents folder covers agents that run in Anthropic’s own sandbox rather than in your process. Sessions there keep files and tool state between turns. The agent SDK folder goes as far as Docker, Kubernetes and Modal hosting recipes.
The one-notebook folders carry the freshest ideas. You’ll miss them unless you scroll the whole file listing.
The table of contents is out of date
The README’s table of recipes covers classification, retrieval, summarising, tool use, third-party retrieval, embeddings, vision, sub-agents, PDF uploads, evaluations, JSON mode, moderation, and prompt caching. That’s a fine list for 2024. It never mentions managed agents, the agent SDK, skills, patterns, observability, extended thinking, or evals. Those folders are most of the repo.
Every link in that table points at github.com/anthropics/anthropic-cookbook, the repo’s old name. They work only because GitHub redirects renamed repos. The contributing guide
still tells you to clone that old path too.
This happens to every recipe repo. Someone writes the contents file by hand while the notebooks arrive through pull requests, so the two drift apart. The same pressure shows up in the 293 open issues and pull requests.
The material is current and the index has fallen behind. Open the directory listing instead of the README, and sort by what you are building.
Two notebooks still call retired models
Reading the model name in every notebook turns up a second kind of drift. 43 of the 91 call claude-sonnet-4-6 and another 21 call claude-opus-4-1. Two are stuck much further back.
The Wikipedia search notebook
still asks for claude-2, and the Pinecone retrieval notebook
asks for claude-2.1. Both models are retired, so those two notebooks fail on a fresh run until you swap the string. The fix is a one-line edit, once you know to check.
The repo’s own tooling shows why. Anthropic runs a Claude Code command called model-check. It pulls the live model list and flags retired names. However, it reads only the files changed in a pull request. A notebook nobody has touched since 2023 never enters the check, so its model string ages in place.
Which notebooks are worth your afternoon
Reading the whole repo as a course is a mistake. Pick the job you have and take the code from the one notebook that covers it.
- Your first tool-using agent starts with the customer service agent and calculator tool notebooks, which walk the loop end to end.
- To wire Claude to your own documents, read retrieval augmented generation first, then pick the third-party notebook for whichever vector database you already run.
- The prompt caching notebook cuts the bill on repeated context, and it’s the highest-return recipe here for most apps.
- Building evals and the evals folder tell you whether a prompt actually works. Without one, you are guessing at whether a change helped.
- The managed agents folder covers agents you don’t have to host, and it’s the newest material in the repo.
- For documents and images, the multimodal folder has reading charts and graphs and transcribing text from forms .
- The JSON mode notebook gets you predictable output, which is the problem every integration hits in week one.
The entry point in managed agents is the failing-tests notebook. It introduces the agent, environment and session objects every other file in that folder reuses.
What it costs to run and how to use the code
You need a Claude API key. The notebooks call the real API, so nothing runs offline and every run bills tokens to your account. Agent notebooks cost more than prompt notebooks, because each run loops several times instead of calling once.
Anthropic’s coordinator notebook split one research job across more workers, each with a narrower brief. That raised the bill instead of cutting it. Every worker thread pays a fixed setup cost, so past a point extra workers just add to the invoice. There is also an observability notebook that pulls your real usage and spend from the API. It’s the cheapest way to see what an experiment did to your bill.
A few things to know before you clone:
- The skills notebooks need beta headers on the client (
code-execution,files-apiandskills), so they won’t run against a plain default client. - The managed agents notebooks need no extra beta flag, just the key and the hosted sandbox.
- The contributing guide asks for Python 3.11 or newer, plus
uvor pip and pre-commit hooks. - The examples are Python, but the ideas port to any language with an HTTP client.
The code is MIT licensed, so you can lift a notebook straight into your own project without a legal conversation. Still, treat them as reference examples. They aim for clarity, so you will need to add error handling, retries and rate limit backoff yourself.
If you’re new to the API, start somewhere else first. The README points at Anthropic’s API fundamentals course , a gentler start in a separate repo. Contributions are open too. Check the open issues and pull requests before you add a recipe.