Figma co-founder Dylan Field thinks MCP is not a perfect protocol

We have reached the phase of the AI cycle where everyone is fighting over the plumbing. Anthropic recently introduced the Model Context Protocol (MCP), a standard designed to help AI agents connect to data sources like Google Drive, Slack, and GitHub without custom code for every integration. It sounds great on paper, but the reality of implementation is hitting the enterprise wall.

Dylan Field, the co-founder of Figma, recently voiced what many founders are thinking: MCP is not a perfect protocol. During a recent industry discussion, Field pointed out that while standards are necessary, the current iteration of MCP creates a friction point between accessibility and security. For a company like Figma, which sits on a massive trove of proprietary design data, the stakes are higher than just making an LLM smarter.

The Illusion of Universal Connectivity

The promise of MCP is that it becomes the universal adapter for AI. In the old world of SaaS, if you wanted two apps to talk, you spent months building a bespoke API integration. MCP wants to replace that with a standardized way for an AI model to ask for a file, query a database, or perform an action. For builders, this is the dream. It lowers the barrier to entry for creating agentic workflows.

However, Field’s critique centers on the fact that a protocol is only as good as the data it can safely move. If you are building a tool where the primary value is the intellectual property contained within the canvas, you can’t just open the firehose. Field noted that Figma intends to support as many standards as possible, but they aren’t going to expose all of their data just to satisfy a new trend.

Why Builders Should Be Skeptical of One-Size-Fits-All

If you are a founder building in the AI space, you need to look past the marketing of “open protocols.” Anthropic wants MCP to win because it makes their models more useful by default. But for the platforms providing the data, the incentives are different. Figma, Salesforce, and Microsoft are all building their own internal agents. They have no reason to make it easy for a third-party Claude instance to scrape their users' data and provide a better experience than their own native tools.

Field’s comments highlight a technical reality: MCP is still early. It lacks the nuanced permissioning layers that enterprise software requires. When a designer at a Fortune 500 company uses an AI agent, that agent shouldn’t necessarily have access to every design file in the organization just because it’s using a standardized protocol. The “context” in Model Context Protocol is currently too broad for companies that live and die by their security audits.

The Platform Moat vs. The Agentic Future

The tension here is about who owns the user relationship. If I can use an MCP-enabled agent to manipulate a Figma file without ever opening Figma, the value of Figma as a destination starts to erode. Field is a smart operator; he knows that being “protocol-agnostic” is a way of saying, “We will play ball, but only on our terms.”

For builders, this means you shouldn't bet your entire architecture on a single protocol like MCP. We are currently in the middle of a standards war. Just as we saw with early web protocols or even wireless charging, the first mover (Anthropic) isn’t guaranteed to be the final standard. There is a high probability that we see a competing protocol from OpenAI or a consortium of legacy SaaS providers who want to protect their moats.

The Founder's Perspective on Integration

When I talk to founders, I always tell them to look at where the power is moving. Right now, the power is moving away from the UI and toward the data layer. Field is signaling that Figma will keep a tight grip on that data layer. They will support MCP for basic tasks, but the deep, meaningful interactions will likely remain locked behind their own proprietary systems.

If you are building an AI startup, don’t assume that because a protocol exists, the data will be free. You are still going to have to negotiate with the gatekeepers. The lesson from Figma’s stance is that the biggest players in the industry are not ready to give up control. They are going to use these protocols as marketing check-boxes rather than true open gates.

“We want to support standards, but we have to be careful about what we expose.” — This is the mantra of every successful SaaS company in the AI era.

What This Means for the Next 12 Months

Expect a lot of noise about “open AI standards.” Don’t let it distract you. The real work is happening in the middle layer—the software that manages the permissions, the filtering, and the sanitization of data before it ever hits a protocol like MCP. Field is right to be skeptical because he understands that in enterprise software, a “perfect protocol” is one that balances utility with extreme caution.

Builders should focus on creating systems that can swap protocols easily. Don’t hard-code your AI agents to only understand MCP. Build an abstraction layer that allows you to plug into whatever Figma, Google, or Microsoft decides is the standard of the month. The goal is to be the bridge, not the architect of the bridge.

Takeaway for Builders

Don’t wait for a perfect protocol to start building, but don’t trust that any one protocol will solve your data access problems. The Figma stance proves that the giants will protect their data moats at all costs. Your job as a founder is to build around those moats, providing value that the platforms themselves are too slow or too protective to offer.

Read the original at TechCrunch AI →