Manufact's $6.3M Seed Round Bets on MCP as AI's Middleware Layer
Manufact, the YC-backed startup formerly known as mcp-use, just closed a $6.3 million seed round led by Peak XV to build infrastructure around the Model Context Protocol. The round itself is modest by 2026 standards, where agentic AI startups routinely close nine-figure deals. But the bet Manufact is making, that the critical chokepoint in AI isn't the models themselves but the connective tissue between models and the systems they need to act on, deserves more attention than the check size suggests.
The company's thesis is straightforward: as AI agents move from demos to production, someone needs to own the plumbing. MCP, the open standard Anthropic released in late 2024 to connect large language models to external tools and data sources, has rapidly become the default integration protocol across the industry. OpenAI, Google, Microsoft, and Amazon have all adopted it. Manufact wants to be the company that makes MCP actually work at scale. The question is whether that's a billion-dollar business or a feature that gets absorbed by the platforms above it.
The Protocol That Ate AI Integration
To understand why Manufact matters, you need to understand the speed at which MCP consolidated a fragmented landscape. Before MCP, every AI application that needed to interact with external systems, reading a database, calling an API, browsing files, required custom integration code. Each model provider had its own function-calling format. Each tool required bespoke wrappers. The result was a mess of incompatible glue code that made production AI deployments brittle and expensive to maintain.
Anthropic released MCP in November 2024 as an open protocol to standardize these connections. The timing was shrewd. The industry was desperate for a shared standard, and Anthropic moved first with something simple enough to adopt quickly but extensible enough to handle real workloads. Within months, the TypeScript SDK accumulated over 34,700 dependent projects. By early 2026, the major model providers had all committed to supporting it.
The analogy Manufact's backers use is "USB-C for AI," and it's apt in ways they may not intend. USB-C succeeded not because it was technically superior to every alternative, but because the industry collectively decided that the cost of fragmentation exceeded the cost of standardization. MCP is winning the same way. Google's competing Agent2Agent protocol and the ADK Task API address overlapping problems but haven't generated the same gravitational pull. The network effects are compounding: more MCP servers means more reasons to build MCP clients, which means more MCP servers. Manufact is positioning itself at the center of that flywheel.
The Middleware Opportunity Nobody Wants to Talk About
Here's the uncomfortable truth about MCP in production: the protocol is elegant, but deploying it is not. Running MCP servers at enterprise scale requires handling authentication, rate limiting, server orchestration, monitoring, and security policies that the open protocol spec doesn't address. This is the gap Manufact is targeting.
The company started as an open-source library that let developers connect any LLM to any MCP server in about six lines of code. That was the hook. The business model is the hosted platform that handles everything the six lines don't: deployment, scaling, multi-server orchestration, and enterprise security controls. It's the classic open-source-to-cloud playbook that built companies like HashiCorp and Elastic.
Manufact claims that 20% of US Fortune 500 companies are already using its tools. NVIDIA, NASA, and SAP are cited as enterprise adopters. The SDK has crossed 100,000 downloads and nearly 5,000 GitHub stars. These are respectable numbers for a company that's barely a year old, but they also raise a question: if enterprises are already building on MCP, why do they need Manufact specifically?
The answer lies in a pattern that repeats across every infrastructure wave. When a protocol becomes standard, the value migrates from "can you use it" to "can you operate it reliably." HTTP became standard; Nginx and then Cloudflare built enormous businesses on making HTTP work at scale. Kubernetes became standard; Datadog and HashiCorp built businesses on making Kubernetes manageable. MCP is following the same trajectory. The protocol is free and open. The operational layer around it is where the money is.
Who Manufact Is Really Competing Against
The competitive landscape is more complex than Manufact's pitch deck suggests. The most dangerous competitors aren't other MCP startups. They're the platform companies that own the layers above and below.
Docker launched its MCP Gateway in early 2026, offering secure, containerized MCP server management integrated directly into the Docker ecosystem that most development teams already use. This is exactly the kind of infrastructure-adjacent play that can commoditize a startup's value proposition overnight. When Docker says "we'll handle your MCP servers," that's a conversation-ender for a lot of engineering teams.
LangChain and LangGraph have built MCP adapters that let developers treat MCP servers as tools within their existing agent frameworks. For teams already invested in the LangChain ecosystem, there's little incentive to add another vendor to the stack. The integration is good enough, and "good enough" kills startups.
Then there are the model providers themselves. Anthropic, which created MCP, has every incentive to make MCP server management seamless within Claude's ecosystem. OpenAI's Agents SDK already includes native MCP support. Google's ADK handles similar orchestration. These companies have the distribution, the developer relationships, and the economic incentive to absorb MCP tooling into their platforms.
Manufact's defense against all of this is vendor neutrality. The pitch is that enterprises don't want to be locked into a single model provider's MCP implementation. They want a layer that works across Anthropic, OpenAI, Google, and open-source models. That's a real value proposition, but it's the same argument every middleware company has made, and it only works if you can stay ahead of the platforms on features while building enough switching costs to prevent commoditization.
What the $6.3M Actually Buys
The investor list is instructive. Peak XV, the rebranded Sequoia India/Southeast Asia fund, is leading. Liquid 2 Ventures, Joe Montana's fund, is in. Pioneer Fund and YC participated. The Supabase co-founder is an angel. This is a developer-tools investor syndicate, not a generalist AI fund chasing the latest transformer architecture.
That matters because it signals what kind of company Manufact is trying to build. This isn't a model company or an application company. It's an infrastructure company that needs to win on developer experience, reliability, and ecosystem breadth. The $6.3M is seed money to do three things: hire engineers to expand the platform, build enterprise features like audit logging, role-based access, and compliance controls, and establish partnerships with the major model providers to ensure Manufact stays in the critical path.
The challenge is speed. In the current AI infrastructure market, where average funding rounds for agentic AI startups hit $155 million in late 2025, a $6.3 million seed round gives Manufact perhaps 18 months of runway with a team of 10 to 15 engineers. That's enough time to prove product-market fit but not enough to build a moat against well-funded competitors or platform companies that decide to build rather than buy.
Compare this to the broader market: Replit closed $400 million in January 2026 at a $9 billion valuation. Coding and developer-tool startups collectively raised over $3 billion. Manufact is playing in the same ecosystem with two orders of magnitude less capital. That's either disciplined capital efficiency or a funding gap that will need to close quickly.
The Real Question: Is MCP Infrastructure a Product or a Feature?
This is the existential question for Manufact and every company building in the MCP ecosystem. The history of developer infrastructure is littered with companies that built excellent tools around open protocols only to watch that functionality get absorbed into larger platforms. Configuration management became a feature of cloud providers. Container orchestration became a managed service. API gateways became a commodity.
The counterargument is that MCP is different because it sits at a uniquely complex intersection. Managing MCP servers at scale requires handling model-specific quirks, tool versioning, security policies, and multi-tenant isolation in ways that don't map cleanly onto existing infrastructure primitives. A company that deeply understands these challenges can build compounding expertise that platform companies struggle to replicate because they're spread across too many surface areas.
There's also a timing argument. We're in the early innings of AI agent deployment. Most enterprises are still running proofs of concept, not production workloads. The companies that help enterprises cross the chasm from demo to production, that understand the specific failure modes, security requirements, and operational patterns of MCP at scale, will have a knowledge advantage that's hard to replicate even with more capital.
Manufact's founders seem to understand this. Pietro Zullo and Luigi Pederzani built the original mcp-use library before MCP had broad adoption, betting on the protocol early and building community credibility. Bringing on Enrico Toniato, a former IBM Research tech lead with AI reasoning expertise, as CTO signals a move toward the enterprise reliability story that will ultimately determine whether this company survives.
What Happens Next
Three predictions for the MCP infrastructure market over the next 12 months.
First, consolidation is coming fast. There are too many startups building MCP tooling for the current market to support. By early 2027, at least two of the top five MCP infrastructure companies will be acquired by model providers or cloud platforms looking to vertically integrate. Manufact is either an acquirer, an acquisition target, or a survivor. There's no fourth option.
Second, Google's A2A protocol will fade. Not because it's technically inferior, but because MCP's ecosystem momentum is now self-reinforcing. Google will quietly add robust MCP support to its platforms while deprioritizing A2A as a standalone initiative. The protocol wars are effectively over. MCP won the way TCP/IP won: not by being perfect, but by being everywhere first.
Third, the real money will be in MCP security and governance. As enterprises move AI agents from sandboxes to production, the ability to audit, control, and secure what those agents can access through MCP will become a hard requirement. The company that builds the "identity and access management layer for AI agents" on top of MCP will be worth more than all the orchestration tools combined. Manufact's enterprise customer base gives it a shot at being that company, but only if it pivots aggressively toward governance rather than trying to be a general-purpose MCP platform.
Manufact's $6.3 million bet is small in absolute terms but strategically sound. The founders identified the right layer of the stack at the right time. The question isn't whether MCP infrastructure matters. It clearly does. The question is whether a seed-stage startup can build a defensible business in that layer before the giants above it decide the job is theirs. History says the odds are against Manufact. But history also said the odds were against Datadog, Cloudflare, and every other infrastructure company that found a way to be essential before it could be replaced. The next 18 months will tell us which side of that ledger Manufact lands on.