Building with Claude Code: Inside Notion's AI development workflow
Discover why the future of work depends on AI, and how you can get started.
View the instructions for connecting to Notion MCP so you can use Notion & Claude Code together.
Want to see what Notion could look like at your organization? Set up time with our sales team.
<aside> ⌛
Want to explore a Notion workspace at your own pace?
Here’s our interactive tour of Notion → Using Notion to connect knowledge
</aside>
Summary
Notes
Transcript
Welcome everyone to today's webinar, Building with Claude Code inside Notion's AI Development Workflow. I'm Wyatt Horan from Anthropix Applied AI team. I have the lucky job of working with teams at Notion who are building at the frontier of what's possible in AI. My work with customers like Notion on the call today help inform how we continue to improve our models and product layers. engineers find friction points or unlock new patterns, that insight flows directly into our API product roadmap.
Partnerships like the one you'll see today between Notion and Anthropic make Clod better for everyone. Notion was actually one of the earliest Clod code design partners. And today you're going to hear about how their early collaboration with us shaped the product of Clod code into how they use it across their teams. So I genuinely love my job and I'm also excited to introduce you to Notion. Um, One housekeeping note here, recording for this will go live in 24 hours.
There's a question widget on the right, so use it throughout the webinar. We've got 15 minutes at the end for Q&A. And then there's a survey widget. We'd love to hear your feedback. We're iterating pretty quickly, and we want to hear what you want that influences what we showcase in future webinars. So now I'm going to introduce the Notion team. I've worked with them for the past year, first as a design partner, Claude Code, but then they've also scaled our API to be available in their product and across their engineering org.
So I'll let them introduce themselves. Ben, do you want to kick it off?
Yeah, I'm Ben. I do internal developer tools and developer experience at Notion. MJ, you want to go next? Sure.
Hi, I'm MJ. I'm a PM at Notion. I work on our APIs and MCP, and I use Claude Code for prototyping. Sure, Bess? Sure.
I'm Shahrabesh. I'm an engineer on our AI product team, and I use Claude Code for basically day-to-day development. You'll see a lot more about that later.
Awesome to hear. Thanks for introducing yourselves. Here's our roadmap for the next hour. I'm going to share a bit of Anthropic's perspective on where agents are heading, what's new with Opus 4.5, and then I'll mostly hand it over to the Notion team. I want them to talk through setting the stage, Ben will do Notion's Gen AI infrastructure. MJ is going to demo how Claude Code and Notion work together with MCP.
And then Shahrabesh is going to show us how Claude Code is used internally at Notion. some good insights for most of our participants here. And then we've got some Q&A. So if you want some questions answered, please ask that there. I'm figuring out how to do my slides here. Here we go. Okay, so we're going to talk about building agents with Claude. One, this is a level set on Anthropic. Our mission is to ensure transformative AI helps people and society flourish.
So we're doing that today by building frontier models and systems and then deploying them responsibly and then being very rigorous about how we study how those models are used in the real world. Expect AI is going to continue to do real work in the economy and one of the first ways we're seeing that is increasingly intelligent AI empowering agents On the right you'll see in our accelerating pace of development eight major model releases in the last two years. We just shipped over
Let's talk about agentic models. It's a way of saying how capable is the model of doing autonomous work. So Opus 4.5 is our most intelligent model and it's intelligent because it has, it's built for complex reasoning, does analysis and high stakes tasks where you need the model to really think. But then we've also enabled it to be much more token efficient. So that means that our customers can now experience frontier that's much more accessible on a better price point, which enables them to use this model in more agent-accused cases and also in Claude Code, which you'll hear about in a second.
We're continually seeing agents like Claude Code perform many hours of work independently. The agents that our customers build, like Notion, on top of our API, are getting better and better at performing work as a virtual collaborator alongside humans. And we expect agents powered by AI to become increasingly independent and capable. And that's why we're making Claude Code accessible from anywhere. Here we go.
I went past two slides.
music Sorry, can you turn off the media? Okay.
Just to the Notion team, can we see the "Call Claude Anywhere" slide here? Great. Okay, cool. I was hearing some audio in the background. So essentially, we've built these agents to be powered by AI to be increasingly independent and capable. And that's why we're making Claude accessible from anywhere. So it's a model that you can call from increasing numbers of surfaces. At the bottom, we've got Claude as a model foundation.
So Sonnet, Haiku, Opus. And then we've added capabilities on top of those reasoning models. Models should have the ability to remember, to create memories as needed. They should be able to do web search to perform and find information outside of their knowledge training. They should be able to do research where they can look across more information that would saturate one context window. They should be able to orchestrate multiple agents beneath them to do very complex tasks.
They should be able to navigate different file systems to find the right information. and use tools and information from other systems. And then increasingly now, they should be able to use skills, which I'll talk about in a second. Skills are essentially the ability for Claude to be almost fine-tuned in a specific task. So on top of those capabilities, then you have platforms where you can then consume that type of agentic knowledge.
So you have Claude Code for developers, you have Claude AI, which is a chatbot for interaction, and then the Claude developer, platform is where you can access the API and then basically bake that knowledge inside of your own product. What Notion is going to show you today is what happens when you connect these pieces really thoughtfully. They're using Claude code internally for their developers. They've built MCP integration so Claude can work with Notion directly.
And then they're exposing AI capabilities to their end customers via our API. It's the same underlying intelligence, but it's deployed across different surfaces. different users. This is where the industry is heading. It's not going to be a single purpose AI, but intelligence that meets you where you work. And then I do want to show you this. You're probably going to hear an audio maybe come over, but that's okay.
Skills, I think, are really important to focus on right now. We've just released them and it transforms how you work with Claude code. A skill is essentially a folder containing instructions. Maybe it could have code snippets, that teach Claude how to do a specific task well. The core is a skill MD file with just best practices and workflows. And the reason that they're important is that a user can just automatically invoke using natural language a skill.
And so Claude can detect when a skill is relevant for a given task or prompt and then load that skill. Users don't need to use a custom slash command. They don't need to point to a specific MCP and make sure it's enabled. They just describe what and Claude knows how to navigate to find the right skill. It's different than MCPs. So MCPs can give Claude the ability to use tools, but skills usually comprise MCPs.
So it's a natural language call that then goes and acts to get the right integrations. Skills can usually invoke those MCPs. And then the third thing that's really important and valuable here is that these skills are available in a marketplace. create skills that can then make Claude better at Notion types of tasks. And that's all accessible to your Claude Code when your Claude Code has access to this marketplace.
I'm going to skip over the two slides here, so hopefully the audio doesn't show. Here we go. We've built out an increase. Okay, wait, are we seeing the... skill function here? If so, can we just skip to the next slide? Awesome. So we've built out a suite of full capabilities for our agents. I mentioned earlier there's web search and memory, but then now as these models are getting better at actually growing past their context window, we've added context editing, code execution, so that there's the non-determinism aspect of these models is no longer present.
execute code, and then citations we've also had for quite a while. All of these capabilities we've added because they're supporting teams like Notion that actually needs to ship these agents in production. One also important thing that you'll see here is we're going to talk a lot about Claude Code, which is essentially the first agent we did to go develop things as its core agentic capability. But all of the capabilities of that Claude code harness, is what we call it, are also extractable for a lot of other tasks that are beyond just coding.
And we've made that available as a Claude agent SDK that you can make as your own agent that goes and does and performs various tasks. So all of the power that you get in using Claude code exposed in your products to be agentic and spawning various sub-agents, and then understanding what is the task to be done and going off and doing it correctly. One quick note on where I think the future of agents are going.