Naptha’s drive towards the “Internet of Agents” 🌐

CV Deep Dive

Today, we’re talking with Richard Blythman, Co-Founder of Naptha.

Naptha is an emerging player in the agent infrastructure space, founded by Richard to help developers build, deploy, and scale agents on the next generation of cloud platforms. Originally inspired by Richard’s background in generative AI and workflow orchestration, Naptha has quickly grown into one of the most active teams in the MCP ecosystem.

Naptha’s core value proposition is simple: to provide the best developer experience for deploying agent servers—with tooling that makes building on protocols like MCP and A2A fast, production-ready, and accessible to startups and developers without requiring heavy infrastructure work.

Today, Naptha powers the infrastructure behind the emerging "Internet of Agents," helping developers spin up, manage, and scale agents across distributed environments. Their platform includes AutoMCP, backend orchestration layers, an MCP gateway, and soon, a server registry hub—all built with the same spirit of iterative, developer-first product design. With a small and highly technical team, Naptha is shaping how agents will be deployed, connected, and scaled across hybrid cloud environments in the years ahead.

In this conversation, Richard shares how Naptha came to life, why he believes multi-agent scaling is the next frontier for AI infrastructure, and how his team is preparing for a world where agents interact across thousands of different servers and protocols in real time.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Richard 💬

Richard, welcome to Cerebral Valley! First off, introduce yourself and give us a bit of background on you and Naptha. What led you to co-found Naptha?

My background originally was in academia—I did a PhD and a couple of postdocs focused on AI. After that, I worked in big tech for a while as a machine learning engineer, mostly on generative AI projects. I was building digital avatars—capturing audio from a human, syncing it to a generated video of a digital human, and generating gestures. I always gravitated toward workflow orchestration. I spent a lot of time working with frameworks like Kubeflow Pipelines to manage the complexity of products like digital avatars, which require a ton of models and orchestration. That became my core expertise in the big tech world.

Alongside that, I was always trying to build startups. I'd say Naptha is my third or fourth. The first one was focused on human pose estimation—analyzing athletic motion from video for physiotherapy use cases. The second was kind of like a hub, similar to Hugging Face but designed for edge models and decentralized infrastructure.

The inspiration for Naptha came from my obsession with workflow orchestration. When LLM apps started gaining traction, I realized the way we orchestrate applications was going to change completely. Traditional workflows and pipelines just aren’t the same as orchestrating LLM apps. I got into the space super early—right around the time LangChain was emerging.

I believe we were one of the first to deploy a LangChain workflow, as far as I know. I got really good at LLM ops-style orchestration—built a few different product experiments around that. 

MCP is a huge part of your current work - tell us a little bit about how the protocol has shaped the trajectory of Naptha in recent months?

Once MCP came out and we started to see all these protocols start to form, we got really into that space. With Naptha, we were actually using something like MCP over a year ago. We had a proof of concept for a multi-agent system where you had three different agents and an orchestrator, each running on different servers and coordinating across the network.

If you saw Anthropic’s workshop from the AI Engineer Summit, they had a very similar picture of where they want MCP to go. We were working on something like that a year ago—not exactly the same, but close. We were also working on interoperable agent frameworks. Some of those are starting to emerge now, like mcp-agent and fast-agent, where you can build agents using separate MCP servers, meaning the tools can be written in different languages or frameworks. We had something similar about a year ago. It worked not just with interoperable tools but also interoperable agents. We had something like MCP servers—we called them Naptha modules. So we had a lot of similar parts of the stack to MCP, but before MCP existed.

Since MCP came out, we've been rebuilding our stack from scratch to be fully MCP compatible, using our experience from building something similar to create better products for developers. Working with MCP has been amazing.

How would you describe Naptha to an AI engineer less familiar with it? 

There’s the full vision of Naptha, and then there’s the go-to-market side, which is more about shipping product experiments. When people ask what Naptha is working on right now, I tell them we’re building developer tooling to make it easier to build on this new kind of platform we’ve been calling the Internet of Agents, which is made up of a stack of new protocols for agents. That’s the broader vision. The products we’re building today are focused on making it easier for developers to build on top of this new ecosystem that’s emerging around protocols like MCP and A2A.

Tell us about your key users today? Who would you say is finding the most value in what your team is building with Naptha? 

We’re definitely focused on developers. The MCP space is super competitive now—big tech players are getting into it, building MCP meshes on top of Kubernetes and things like that. That’s what Google is doing. I'd say a lot of the big players are going to focus on enterprise, but we want to stay very developer-focused and provide a great developer experience. In that sense, we’re trying to be more like Vercel for MCP. 

It’s inspired by platforms like Vercel, Heroku, and even GitHub Codespaces. For example, Heroku uses a Procfile, which acts as an entry point for applications—and that’s similar to what AutoMCP does. It creates something like a Procfile for MCP servers. Users can add AutoMCP to their existing agent projects, it sets up the entry point, you make a few small edits, and then you can easily deploy on top of our platform. None of the other solutions right now make it that easy from a developer point of view, and that’s what we want to compete on.

Walk us through Naptha’s platform. What use-case should developers experiment with first, and how easy is it for them to get started? 

On the backend side, Naptha packages everything you need to run agents on the cloud or your laptop, and for those agents to interact with other agents on the network. That includes servers with various communication protocols, orchestration, local inference and storage. V1 is fully open source but wasn’t based on MCP servers. V2 is fully MCP compatible, and you can deploy your agents on it via our hosted platform, but it’s not yet open source.

 Then we’re working on tooling that makes it easy to deploy on top of that backend. That’s what AutoMCP is—it’s a very simple library that generates the Procfile, which we call run_mcp.py, for your existing agent project.

Plus, we're about to launch a hub—a registry for agent and tool servers—in the next one to two weeks. 

On the agent framework side, we built the first interoperable agent framework where you can use, for example, a CrewAI agent with a LangChain agent with a typescript agent in the same orchestration flow. It was based on modules, our version of MCP servers. These days you have mcp-agent and fast-agent, which do something similar for interoperable tools and working towards interoperable agents. I recommend you check out those. We’re just happy other teams are starting to build on a similar stack, and we can focus on the platform.

There’s so much of the stack to build—this whole platform—and honestly, we’ve been guilty of building too much at once over the past year. When you're offering so many different things, it makes it harder to explain what you’re actually offering. Even a year ago, it was really difficult to explain the concept of the Internet of Agents to people. I'd say now, with MCP becoming more understood, it’s a bit easier to communicate that vision of a new platform. But our answer to that has been to just build one thing at a time that provides value to users on its own.

When Cloudflare launched, they first rolled out their CDN—the content delivery network. Later, they launched Cloudflare Workers, and after that, a competitor to S3 storage. They’re building an entirely new platform too, but they did it iteratively. I'd say Naptha’s approach will be similar: offering a hosting platform for MCP servers as one product experiment, then separately building the MCP Gateway, which is kind of like the CDN equivalent for MCP. We’re really focused on solving developer pain points one step at a time, without needing the whole platform to exist right away. It’s a big task, but hopefully we can iterate our way there.

Tell us about an existing use-case for Naptha that has delighted you the most. 

I'd say being able to use MCP servers through clients like Cursor has been a big unlock for me as a developer. I’ve built a ton of agents over the past few years, but usually I’d run them through a Python script or a CLI or something like that. We never really wanted to build the interface ourselves. But now that I can access all these agents through Cursor, it’s unlocked a lot.

A lot of the time now, if I’m working on an agent, I can just host it really easily on the AutoMCP platform and interact with it directly through Cursor. Cursor has basically become my main interface for interacting with all the agents I build, which is super useful. It’s unlocked a bunch of use cases too. I use Cursor for more than just coding now—I use it for note-taking and other tasks—and it means I can still interact with agents as MCP servers inside my notes. They can iterate on my notes, help refine them, and so on. I'd say there’s a much broader set of use cases for agent MCP servers through clients like Cursor that MCP has made possible.

The ones I’m particularly excited about are based around multi-agent use cases. For example, I can interact with my coding agent on Cursor, while my coding agent chats with the coding agents of other Naptha team members in a group chat. 

You’ve mentioned an internet of a billion agents working in parallel - how close or far away are we from that reality? What role do you think Naptha will play in bringing that to life? 

My prediction at the start of the year was that we would see workflows with 1 billion agents in 2025. We’ve solved most of the engineering challenges associated with running them, but there are still challenges on the research side. How do you coordinate multi-agent systems at that scale, without handcrafting graphs of interactions and communication? We’re doing research in fields including game theory, mechanism design, and evolutionary approaches to achieve that.

What makes an AI agent most-effective, and how are you measuring that at Naptha?

I'd say context and memory are really important, and while the space has kind of solved these through frameworks like LangChain—people have been using memory with LangChain for a while—I haven’t seen many solutions that combine memory and RAG with MCP yet. That’s still super nascent. I’d love to see more agents with memory running through MCP, but honestly, that’s still beyond the cutting edge right now.

At least from where we’re seeing it today, the biggest unlock with MCP has been around giving agents access to a lot of tools and handling the routing between them. There are still limits though. Like, can you just add 2,000 MCP servers to your Cursor client and have it work? The answer is kind of—but also not really. So right now, the major value is in tool access and routing. I'd say as MCP matures and figures out how to handle things like resources better, we'll start to see memory and more advanced capabilities come online.

MCP has become a hyper-competitive space. What would you say sets Naptha apart from other players active in the space? 

It's super competitive and getting more competitive all the time. In our first experiments, where we're trying to compete is really on developer experience, because we don't see great developer experience anywhere else right now. There's a risk with that though—Cloudflare is already moving into MCP, and if they really lean into it and offer a developer experience similar to Vercel, it’s going to be hard for us to compete.

If that happens, I'd say we’ll need to fall back on a different way of competing. Ultimately, the platform Naptha envisions is different from the current generation of cloud platforms. You had AWS as the first big cloud platform, and then Cloudflare came along, disrupting AWS by being more edge-focused, with more data center locations.

I'd say we're going to see a new generation of platform emerge—something like a hybrid cloud, or what I sometimes call the Internet of Agents. It’s about agents running everywhere: not just in data centers, but locally, and across different cloud providers like AWS. That's the platform we see for agents—multi-cloud, hybrid cloud.

So if we can't win purely on developer experience, we'll compete by pushing toward this new platform. For example, solving problems like: how can you run an agent locally on a laptop behind a router, and still have that agent interact with others across the network? There’s a lot we’re good at there, and that’s where we can stand out.

Talk to us about some of the research innovations that have made Naptha possible. 

The main research direction Naptha focuses on is multi-agent scaling, which is something we haven’t seen much of yet. There have been different generations of scaling—first, scaling model training, then scaling laws for inference-time compute—but all of that was for a single agent or a single model. What I expect to start emerging, probably next year, are scaling laws for multi-agent systems. We've been closely monitoring research in this direction.

How does it differ from inference-time scaling? You're still running inference cycles, but each inference cycle could use a different model, for example. There’s a lot of research showing that using diverse models with agents—like eight agents with eight different models—is more performant than eight agents all using the same model.

That’s where Naptha is putting in work: funding researchers, contributing to research for conferences like ICLR, and staying close to where the field is moving. It’s directly relevant to the Internet of Agents too. As we scale the number of agent MCP servers on our registry, the next question becomes: how can you run a million agents in parallel and aggregate the outputs in an intelligent way?

I'd say that's going to be a huge unlock, and future improvements in performance are going to depend more on multi-agent scaling than just inference-time optimization.

We’ve touched on MCP already, but tell us about where you think the space is heading. Do you have a horse in the race protocol-wise? 

To a certain extent, we’re pretty agnostic about which protocol ultimately wins. One of the product experiments we’re working on—where we already have a proof of concept—is making it really easy to deploy an agent as an A2A server. That’s something we’re considering launching. There are also other emerging protocols, like AGNTCY from Cisco, among others.

The protocol wars are definitely heating up, but we don’t want to pick sides. Our goal is to provide developers with great tools so they can deploy agents using whatever protocol they prefer. Over time, we’ll see which standards gain the most traction.

Right now, MCP clearly has the most community participation and momentum. A2A was a smart move by Google, and while there’s no direct overlap in the specs between MCP and A2A, there’s definitely some overlap in vision. It remains to be seen whether MCP will eventually incorporate some of the features A2A already includes.

At the end of the day, we’re focused on making developers happy and enabling them to build with whichever protocols best fit their needs. It’s very possible multiple protocols will coexist and serve different use cases.

How do you see Naptha evolving over the next 6-12 months? Any specific developments that developers should be excited about? 

I'd say Naptha already has the best developer experience for deploying MCP servers, which is why people should check it out now. We have a lot of plans in the works to keep improving that experience over time, especially for developers and early-stage startups rather than enterprise. For any startup that wants to host an MCP server in a production environment, we want Naptha to be the go-to platform.

Right now, we’re seeing a lot of proof-of-concept projects where people interact with hosted MCP servers through clients like Cursor, but we haven’t really seen anyone offering true production-grade MCP server hosting for startups. That’s where we’re focused. It's not just about hosting; it also involves things like monitoring, payments, and all the other infrastructure pieces that startups will need. These are all features we’re planning to roll out to support our user base, and that’s what we’ll continue iterating on throughout the year.

As for where I see the space going: the main battle now seems to be shifting toward the platform layer. It used to be all about which model was coming out next, but now models  are starting to become commoditized. GPT-4.5 wasn’t a huge leap compared to GPT-4, and there’s just less excitement around new models in general.

After that, we saw a wave of competition around agent frameworks—every major model provider started launching their own. We’re still seeing a bit of that, but the next big battleground is clearly the platform: the protocol wars, deciding which standards are going to win out.

That’s going to dominate the conversation for the rest of this year. It’ll be fascinating to watch because once the platform layer settles a bit, we’ll start seeing products emerge on top of it. Right now, we don’t even know exactly what those products will look like. What does the Vercel for MCP servers look like? What does memory infrastructure for MCP servers look like? What sort of new MCP clients will we see (Cursor for X)? These are all products we haven’t seen yet, but I'd say we’ll start to see them emerge over the course of 2025.

Lastly, how would you describe the team at Naptha? Are you hiring, and what do you look for in prospective team members joining Naptha? 

We’re about 10 people on the team now, and it’s an awesome group. In the previous phase, we were very focused on R&D—we had some really strong researchers and multi-agent systems builders, and we still have that expertise on the team. But now we’ve started expanding on the product and go-to-market side. We’ve brought in a bunch of former founders, which has been a big focus for us.

Overall, it’s a really strong team, both on the technical and infrastructure side, and now we’re building out the product and GTM side as well. The goal is to structure Naptha in a way that not only solves the hard technical challenges but also consistently ships products that are really useful for users. That’s our focus with the team structure at the moment.

If you’re interested in solving engineering or research problems to run a billion agents across a network of servers, or how to build products on top of the Internet of Agents, we would love to chat to you.

Anything else you'd like our readers to know about the work you’re doing at Naptha?

We’re huge proponents of open source. Beyond the technical and research challenges of scaling multi-agent AGI lies a deeper societal question: who will own the platforms for running AI? At present, the answer tilts heavily toward a small number of firms with the scale and capital to build closed, vertically integrated systems. Protocols like MCP and platforms for the Internet of Agents offer a way to rebalance the equation, opening the field to startups, open-source contributors and lone developers who would otherwise be left out.

We want to build a community around this vision to solve the associated challenges. We’re super active in the MCP community—we're starting to take part in the MCP Hosting Community Working Group, and we're really close with a lot of other people building agent frameworks in the MCP space, like mcp-agent and fast-agent.

We’re hosting a meetup here at ICLR for people interested in agent protocols, and we’re planning to head out to San Francisco to run hackathons around this as well. For anyone who wants to join a community that's really active in the MCP space, we'd love to chat and let you know where we hang out. 

Conclusion

Stay up to date on the latest with Naptha, learn more about them here.

Read our past few Deep Dives below:

• V7 Go - AI Agents for Complex Document Workflows 🎛

• Zencoder’s Repo Grokking - Your AI coding unlock 🔑

• Rime - Authentic AI Voices for Real-time Conversations at Scale 🗣️

• Future AGI - Built for End-to-End AI Development 💡

• Antithesis - the last word in autonomous software testing 🎛

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs