The MCP Gold Rush
The tech world has a new obsession: Model Context Protocol (MCP).
MCP, introduced by Anthropic in November 2024, standardizes how AI assistants connect to external tools - like USB-C for AI applications. This allows AI assistants like ChatGPT or Claude to connect to external applications like Linear, Github, and Figma.
For the past several months, developer Twitter has been buzzing with MCP tutorials and success stories. Every AI startup seems to be scrambling to add MCP integration to their roadmap. The message seems clear: MCP is the future of AI tool integration.
But we don't think that MCP is necessarily the right architecture for all AI apps. The core limitation, in our view, is that MCP treats the client application as separate from the MCP server.
So for example:
ChatGPT is the client, Linear is the server
Claude is the client, Github is the server
Cursor is the client, Figma is the server
In every case, two different apps are involved - the client, which is an AI assistant, and the server, which is a traditional application. At Quotient, we often refer to these as Software 2.0 and Software 1.0 applications, respectively.
MCP integrations like these are a great way to give generic AI assistants access to external tools, and it makes sense for some use cases.
But we deliberately decided not to build Quotient this way, because we want the client and the server to be part of the same application.
Integrating Client and Server
Our goal at Quotient has always been to give our customers marketing superpowers with the help of AI. To that end, we built a platform that tightly integrates our AI agents (the “client”) with traditional marketing software tools (the “server”).
An alternative way to solve this problem would be to connect generic MCP clients (like ChatGPT) to a bunch of marketing tools (like Hubspot) via MCP. But we believe this would be a clunky experience, at best.
Why is it clunky? For starters, it means you have to use two different tools to get the job done - your chat application and your marketing platform.
That means:
Two different tabs open in your browser
Paying for two different subscriptions
Two different logins to manage
This alone is cumbersome for users.
But more importantly, this separation dramatically limits the depth of integration between the AI agents and the software 1.0 systems they use.
With Quotient, we wanted a much tighter integration between our AI systems and the rest of our marketing stack. Specifically, we had three goals for the product that could only be achieved through tight integration between our AI agents and our marketing stack.
1. Side by Side Collaboration in the UI
We wanted Quotient’s agents to be able to collaborate with users side-by-side in our user interface. We wanted the agents to understand what page the user was on, what part of the UI they were pointing to. When a user made edits, we wanted the agent to “see” those edits instantaneously and take them into account.
Conversely, we wanted the user to be able to see the agent’s changes immediately. For example, when an agent adds a paragraph to a blog, or tweaks the styling of an email broadcast, we want the user to see that immediately on their screen.
This type of thing is not really possible with a traditional MCP integration. To accomplish this, we needed to build a multiplayer editing system and user interface that allowed humans and agents to collaborate in realtime on various artifacts.
This leads us to our next design goal.
2. Robust Artifact Editing
Most APIs for external systems are highly limited. They usually have some basic create/read/update/delete functionality but they typically don't allow an agent to do sophisticated tasks like:
Doing targeted visual updates on an email broadcast (e.g. updating CSS, replacing text, or saving blocks of HTML as reusable components)
Updating and reformatting specific passages of a blog
Writing a sophisticated query against database of people and companies
Most MCP servers do not allow for this level of granularity, because most of MCP servers were not actually built with agents in mind. For the most part, they just duct tape tools on top of existing APIs.
At Quotient, we designed our data structures for things like blogs and emails to be agent-friendly from day one. We created data structures that agents can easily manipulate with tools, creating a much more seamless experience compared to MCP.
3. Comprehensive Context
The ability to share context over MCP is highly limited. You can share some basic snippets but that's about it. At Quotient, we take context engineering extremely seriously, and we have sophisticated systems for giving agents exactly the data that they need to perform the task at hand.
For example, all of our agents receive access to each brand's Knowledge Store — a set of curated brand guidelines for each Quotient business. Agents also receive additional context about the state of the Quotient account, e.g. what marketing campaigns are currently going on, what automated workflows are running, and more.
This degree of context engineering is not possible with a traditional MCP integration - it requires building the MCP client and the MCP server in tandem. But this degree of contextual understanding is critical for building agents that truly understand your business the way you’d expect a human colleague to.
A Concrete Example
Let's get concrete about the difference. Imagine you want to launch a product announcement campaign that spans email, blog content, and social media.
With MCP: You'd use ChatGPT to generate content ideas, manually copy-paste between your email platform and blog CMS, coordinate timing across different tools, and manage audiences separately in each system. Each step requires context-switching, manual coordination, and hoping the messaging stays consistent, which is challenging because ChatGPT doesn’t truly understand your brand.
With Quotient: Our Campaign Agent automatically understands your product positioning, coordinates across our Blog, Email, and Web agents to create a cohesive campaign with consistent messaging, proper sequencing, and unified analytics - all automatically. What takes weeks with traditional tools takes minutes with integrated AI agents.
The difference isn't just convenience - it's intelligence. MCP gives you access to tools. Quotient gives you a marketing team that happens to be powered by AI.
Rethinking the Foundation
Most people think MCP is the only way to build integrated AI applications. They assume it's all about giving ChatGPT or Claude access to your existing tools. But that assumes the tools we have today are the right foundation for an AI-powered future.
Instead of retrofitting legacy software with MCP servers, we built new applications from the ground up with AI-native architecture. We designed data structures, user interfaces, and workflows specifically for human-AI collaboration rather than bolting AI onto systems designed for humans alone.
The result isn't just better tool integration. It's a fundamentally different experience. One where AI doesn't feel like an external assistant accessing your tools, but like an intelligent extension of the tools themselves.
MCP will continue to thrive, and that's fine. It serves an important role in the ecosystem. But there's a more interesting path forward—one where we don't just connect AI to our tools, but reimagine what tools can be when AI is part of their DNA from day one.
