Learn and master Clay with Clay university

You'll learn to build a Claygent from a plain-language description, test it, and tune the settings that matter most: the model, the outputs, the context, and the tools.

🌐 Why Claygent Matters

Claygent isn't new, but it has gotten much better. Accurate AI web research is one of Clay's biggest differentiators. Plenty of teams build their own GTM systems outside Clay, and having a reliable research agent on hand, a small computer you can point at any question, is what tends to set Clay apart.

✍️ Building One in Plain Language

Open Claygents in the left sidebar. There's now a dedicated screen for every Claygent in your workspace. Describe what you want, and AI builds it.

Here's the Claygent from the video, a research assistant for personal context on a person:

Build a Claygent that finds personally relevant information about an individual that may not be straightforward to find. Rank-order the research:
1. Interviews, podcasts, blog posts, or any personal media, starting with anything related to their current job
2. Their personality and interests: think concert tickets, books, movies, hobbies
3. Anything related to a previous job, as long as it's weighty and interesting
If nothing turns up in those categories, look for anything else that adds flavor or context about the person.


Why bother? Most B2B data doesn't tell you anything that helps you build a real connection. Once, someone noticed a tweet where Yash mentioned a specific chocolate cookie from India, sent him some, and asked to chat. He took the call. That works for cold email, and it works just as well for existing customers and partners.

🧪 Testing the Prompt

You can use AI to build AI. What you need to be good at is taste: is this a good prompt? And troubleshooting: do the test results hold up?

1. Read the prompt it wrote. Skimming is fine for a demo. For a prompt that matters, read it all.
2. Check the inputs. Here it uses name, company, and title.
3. Run the suggested test record. Clay suggests test data automatically.
4. Switch to a table of tests to see several results at once. These are tests, not a production run.
5. Add harder test cases. Well-known people are easy. Add a few lesser-known people who look like the ones you'll actually run it on.

🎛️ The Settings That Matter

• Model: don't always let AI pick it. If a data point is valuable to you, spend more on it. You probably wouldn't run the most expensive model across a huge list, though.
• Inputs: here, every input is optional, which makes sense.
• Outputs: separate fields make the full result easiest to read. This one came back as a JSON schema instead.
• Context: account context and business context (your AI context) give the Claygent what Clay already knows.
• Web search and tools: which ones are available depends on the model you pick, so pick a model with tool access when tools matter to the prompt.

Tools: The Most Underused Part

You can give a Claygent any other tool in Clay. That means a default enrichment, or one of your Functions. You can even add a custom MCP server, which connects any tool in your stack that has an MCP connection.

For example, connect your call recording tool's MCP, and the Claygent can mine your call history with an existing customer before it writes anything. Not sure how to set one up? Ask your AI agent how to add a custom MCP server in Clay. It will walk you through most of it and tell you the URL and API key you need.

🚀 What's Next

You can now build a Claygent from a plain-language description, test it on real cases, and give it the model, context, and tools it needs.

That's the end of the bonus lessons. Go back to the main play in Lessons 1–5 and add the pieces you just learned: signals on your segments, a Function in your workflow, a Claygent with your call history.