Manufact's Open-Source Library Powers NASA's MCP Agents — Here's How

The Connector Layer Gets Its Backbone

Manufact, the San Francisco startup from Y Combinator's Summer 2025 batch, closed a $6.3 million seed round reported by VentureBeat in March 2026. The raise signals that investors see the Model Context Protocol graduating from open-source experiment to production infrastructure. Peak XV Partners, formerly Sequoia Capital India and Southeast Asia, noted in its Surge 11 announcement that AI is "beginning to transform traditional industries from banking and healthcare, to semiconductors and manufacturing."

The company began as mcp-use, an open-source SDK framework built by Pietro Zullo and Luigi Pederzani. Pederzani previously led a 12-engineer team delivering projects at Accenture Switzerland and served as founding engineer at Morgen, an ETH spin-off used by teams at Spotify, GitHub, Linear, and Canva. They didn't set out to build a cloud platform. They built the SDKs — TypeScript and Python — and the MCP Inspector, watching developers struggle to take local MCP servers to production. "mcp-use is to Manufact what Next.js is to Vercel," the founders say.

Nine million SDK downloads across both languages. Manufact's data shows ten thousand GitHub stars. Open-source from day one. Manufact now sits as the #1 recommended deployment option in OpenAI's own documentation. The platform turns a git push into a live MCP app or server in under 60 seconds, running six categories of publishing checks mapped to store requirements: protocol and discovery, tool conformance, security and policy, metadata and configuration, domain/TLS/CSP, and assets.

Growth signals accelerated after Anthropic shipped MCP UI in Claude. "Since mcp ui was released in claude we got a 2x boost in retention," Manufact reported. Listing on the Claude Connectors marketplace triggered a signup surge that continues. Early YC batchmates AgentMail and Item used Manufact to reach that marketplace. Larger companies in finance, insurance, and marketplace verticals now run production MCP apps on the platform.

Pricing reflects the developer-first motion: a free tier with $5 monthly usage credits, a $25 hobby plan adding preview deployments and end-to-end checks, a $250 startup plan with all regions, and custom enterprise contracts. Metered usage runs $0.10 per 1,000 tool-call requests, $1 per eval run, $2 per E2E check; discovery traffic like tools/list is never billed.

Paul Graham framed the shift bluntly: "AI apps are the new browsers." Customers now do most of their work inside Claude and ChatGPT. The connector marketplaces become a distribution channel: one-click install, instant utility inside the chat interface users already inhabit. But shipping a production-ready MCP server remains its own project: the protocol evolves fast, auth is easy to get wrong, every client behaves differently, and store review can block launches for weeks. Manufact exists to absorb that complexity.

Building the Sales Engine

The seed round didn't just fund engineering headcount — it triggered an immediate push to build a sales engine. Manufact's Y Combinator job board listing for a Sales Development Representative, posted from its San Francisco headquarters with a remote option for U.S. candidates, signals that the six-person team is converting usage traction into pipeline velocity. The role carries a base salary of $60,000 to $100,000 plus 0.05% to 0.20% equity, and explicitly welcomes new graduates (a hiring profile that matches a startup moving from founder-led sales to a repeatable outbound motion).

The timing aligns with what Manufact has disclosed about platform adoption: a surge of signups after the Claude Connectors marketplace launch. That growth curve creates an inflection point. Founders Pietro Zullo and Luigi Pederzani — both technical — can no longer carry the sales load alone. The SDR hire is the first dedicated revenue role, tasked with qualifying inbound interest from the Claude Connectors marketplace and ChatGPT plugin store, where Manufact has already helped YC batchmates AgentMail and Item get listed, and from larger accounts in those verticals running MCP apps in production.

San Francisco remains the anchor. The YC listing specifies the city as the primary location, with remote flexibility only within the United States. That constraint reflects the density of early MCP adopters in the Bay Area: AI labs, agent startups, and enterprise innovation groups all within meeting distance. An SDR sitting in SoMa can walk into partner offices, attend the same hack nights as prospective customers, and feed the founding team real-time feedback on which verticals are converting fastest. The equity band, narrow at 0.05% to 0.20%, is typical for a post-seed company with six employees; it also suggests Manufact expects to hire more than one SDR over the next 12 months, diluting the pool gradually rather than granting outsized grants to a single early rep.

At $60,000 to $100,000 base, Manufact competes with well-capitalized AI infrastructure peers for a talent pool that industry veterans describe as fiercely contested. A breakdown from a 13-year SDR hiring veteran noted that the average rep now makes 40 to 60 dials a day and that hundreds of applicants chase each posting. Manufact's willingness to consider new grads widens the funnel, but the equity component is the differentiator: a bet that MCP becomes the de facto connectivity layer for AI agents, and that early sales hires capture upside from every enterprise deployment.

What the job posting doesn't say — but the usage data implies — is that the inbound queue already outpaces founder capacity. Manufact's platform sits between MCP server developers and the marketplaces where AI agents discover them. Every new server published, every enterprise that moves a pilot to production, generates a conversation that needs qualification, demo scheduling, and handoff to a closing motion the company has yet to formally build. The SDR role is the bridge. Without it, the signup growth rate becomes a backlog rather than a revenue curve.

The hire also positions Manufact for the competitive pressure outlined below: rival MCP cloud platforms are launching, and the race to lock in the default deployment layer will be won partly on distribution speed. A dedicated SDR team, even starting with one person, lets Manufact systematize the outreach that today relies on founder networks and YC batch connections. It turns "we've helped startups get onto Claude Connectors" into a repeatable motion: identify developers building MCP servers, show them the deployment and observability stack, convert them to paid plans before a competitor does.

For candidates, the role is a front-row seat to the standardization of AI agent connectivity. For Manufact, it's the first step toward a sales organization that can match the platform's technical adoption curve. The job is live on the Y Combinator board; the next quarter will reveal whether one SDR suffices or whether the pipeline demands a pod structure sooner than the headcount plan assumes.

Enterprise Logos That Signal Trust

NASA is building an agent with MCP using Manufact's open-source library. The agency's developers confirmed it publicly — a rare, named endorsement in a space where most enterprise pilots stay behind NDAs. When a federal research organization with strict security requirements chooses a protocol for production agent work, the evaluation bar has already been cleared. NASA's own job board shows seven roles added in the past week alone, including a Director, Science Directorate and multiple data scientist positions, according to Zero G Talent's board data; this is evidence that the agency's MCP adoption is expanding, not experimenting.

The logo wall on Manufact's site tells the rest of the story. NVIDIA, IBM, Oracle, Red Hat, Verizon, Intuit, Elastic — each represents a separate procurement review, a separate security questionnaire, a separate decision to standardize on MCP rather than build proprietary agent-to-tool bridges. These are not proof-of-concept trials. The companies listed run MCP apps on Manufact Cloud in production. Finance, insurance, and marketplace firms (sectors that move slowly on new infrastructure) have already shipped.

OpenAI's documentation now lists Manufact as its top recommended deployment option for MCP servers. That placement came without a partnership announcement; the docs simply reflect what OpenAI's own engineers observed in the wild. Anthropic's Claude Connectors marketplace shows the same pattern: Manufact-listed servers surged after the platform appeared there, and the signup wave "is still going on now," per the YC profile.

The protocol itself, Model Context Protocol, functions as the connective tissue. Anthropic introduced it in late 2024 to standardize how LLMs call external tools and data sources. OpenAI adopted it months later. Google DeepMind added support. The result is a single interface that works across Claude, ChatGPT, and custom agents. Manufact's platform sits at the deployment layer: push to GitHub, get a preview deployment per pull request, run cross-client conformance tests automatically, then publish to both marketplaces with generated assets that satisfy reviewer checklists.

Enterprise buyers care about three things the platform addresses: authentication done correctly across clients, observability when tools fail in production, and a path through marketplace review that doesn't stall for weeks. The SDK handles the first. The cloud inspector, session replay, and trace analytics handle the second. The publishing checklists and auto-generated submission assets handle the third.

SAP does not appear in Manufact's published customer logos or case studies. The enterprise roster documented so far clusters around cloud infrastructure (Elastic, Red Hat, Oracle), AI hardware (NVIDIA), financial services (Intuit, Verizon), and government research (NASA). That distribution mirrors where agent workloads are moving first: environments that already run containerized microservices, have mature DevOps pipelines, and need to expose internal APIs to LLM-driven workflows without rewriting them.

The adoption curve resembles Kubernetes circa 2017: an open-source project graduates to managed service, early adopters are the companies that felt the pain most acutely, and the reference implementations become the default choice for the next wave. Manufact's seed round bets that the managed-service moment for MCP has arrived. The enterprise logos suggest the bet is landing.

Rivals Crowd the Infrastructure Layer

Manufact's traction has not gone unnoticed. A cluster of startups and incumbents is now racing to stake claims in the MCP infrastructure layer, each targeting a different slice of the deployment, security, and governance stack. The protocol's rapid adoption — backed by every major model maker including OpenAI, Microsoft, AWS, and Google — has turned what was an open-source spec into a commercial battleground.

The most direct challenger is Runlayer, which exited stealth in November 2025 with $11 million from Khosla Ventures' Keith Rabois and Felicis. The company was founded by third-time founder Andrew Berman (previous companies: Nanit and Vowel, which sold to Zapier). Runlayer's pitch centers on the protocol's original blind spot: security. "The MCP protocol itself doesn't include much security out of the box," the team told TechCrunch, "so many MCP implementations have already been found to be vulnerable in a variety of ways." Their platform combines a gateway for agent identification and access control with threat detection that analyzes every MCP request, observability across all permitted servers, and a permissioning model that maps agent rights to existing identity providers like Okta and Entra. In four months, Runlayer signed dozens of customers, including eight unicorns or public companies such as Gusto, dbt Labs, Instacart, and Opendoor.

Open-source project Obot takes a different angle: an Okta-style catalog of pre-vetted MCP servers that IT teams can approve for internal use. It positions itself as a governance layer rather than a deployment platform, letting enterprises curate what their agents can reach without building custom tooling.

The incumbents are moving too. Cloudflare, Docker, and Wiz have all shipped MCP security products, most of them gateway-style controls that sit between agents and the tools they call. Prefect's FastMCP framework lowers the barrier to building servers, while Arcade and Keycard handle agentic authorization and permissioning; this is plumbing that Manufact also addresses but from a deployment-first stance.

Up the stack, a second wave of infrastructure startups is forming. Braintrust, LangChain, Bigspin.ai, and Judgment Labs are building evaluation harnesses, benchmarking environments, and real-time feedback loops for agentic workflows. DataHub adds data lineage and compliance verification. Arklex, Kiln AI, and Pi Labs are rethinking measurement and feedback for AI-native applications. None of these are Manufact clones, but they all assume MCP as the connective tissue, and they all need a place to run, test, and distribute the servers they help create.

Anthropic itself has entered the fray with the Claude Apps directory, launching integrations for Slack, Canva, Figma, Box, and Clay (Salesforce is noted as imminent). The directory is gated behind paid tiers (Pro, Team, Enterprise) and signals Anthropic's intent to own the application layer atop MCP. That puts it in competition with Manufact's publishing and distribution ambitions, even as the protocol remains open.

The pattern is familiar: a standard wins adoption, then a platform war erupts around operationalizing it. Manufact's seed round and reported signup surge have accelerated that timeline. The question now is whether the market consolidates around a few horizontal platforms (deployment, security, observability, eval) or fragments into vertical-specific stacks.

Where Analysts See the Market Heading

No dedicated market sizing exists yet for the MCP cloud segment; analysts at Deloitte, Gartner, and IDC have not published a standalone forecast. The category is too new. Instead, MCP infrastructure is tracked as a sub-component of the broader AI agent and model-serving layer, which itself sits inside the hyperscaler capex surge now reshaping global compute demand.

The proxy numbers are staggering. The top five hyperscalers (Amazon, Microsoft, Alphabet, Meta, and Oracle) are on track to spend roughly $443 billion on capital expenditures in 2025, with CreditSights projecting that figure to climb to $602 billion in 2026, a 36% year-over-year increase. Analysts estimate roughly three-quarters of that spending will flow directly into AI infrastructure. Morgan Stanley and JPMorgan calculate the AI infrastructure push could drive up to $1.5 trillion in additional borrowing by tech companies in the coming years; UBS sees as much as $900 billion in new issuance in 2026 alone. OpenAI's Stargate projects alone represent roughly $850 billion in headline commitments, nearly half the $2 trillion global AI infrastructure surge HSBC now forecasts.

Data center demand tells the same story. Deloitte projects data center load could reach 176 gigawatts by 2035, a fivefold jump from 2024. Two terawatts of capacity are already stuck in interconnection queues, almost twice the currently installed base. Lead times for transformers and switchgear have stretched to multiple years. The cost of a new gas-fired plant has surged to more than two and a half times projects built just a few years ago. In the first nine months of 2025, M&A in the US electric power sector exceeded $109 billion. Every layer of the stack (power, land, fiber, chips, cooling) is constrained.

Against that backdrop, Manufact's position is defined by timing more than market share. The company emerged from that batch with a reported surge of signups. Enterprise adoption signals are concrete: NASA uses the open-source library, and NVIDIA developers use the open-source tools. That customer set, spanning space exploration and the dominant AI compute vendor, suggests Manufact has secured early footholds in the verticals most likely to standardize on MCP first: government R&D and model providers.

The competitive response validates the signal. Section 4 documents rival startups launching competing MCP cloud services. When well-capitalized entrants appear this early, it typically indicates venture investors see a category-forming moment, not a niche. The risk for Manufact is not demand but execution: the platform must harden its observability, testing, and publishing workflows before better-funded competitors replicate the core loop and outspend on distribution.

Qualitatively, three scenarios frame the next 18 months. In a base case, MCP becomes the de facto connector for agentic workflows across ChatGPT, Claude, and enterprise LLMs, and Manufact captures a low-single-digit slice of the model-serving layer, a market measured in the tens of billions by 2026. In a bull case, MCP evolves into the "USB-C for AI" that the VentureBeat article highlighted, and Manufact's first-mover advantage in developer experience compounds into a durable platform moat. In a bear case, hyperscalers absorb the protocol into their native tooling (AWS Bedrock, Azure AI Studio, Vertex AI) and the independent cloud layer compresses to a feature, not a product.

The hiring signal is the clearest near-term indicator. Manufact's SDR hire in San Francisco reflects a company racing to convert inbound developer interest into enterprise contracts before the window narrows. If Manufact can turn that class of customer into a repeatable motion, the market share question shifts from "what percentage" to "how fast can they onboard the next hundred."

Working in frontier tech? Zero G Talent tracks the openings: see every open NASA role, browse frontier tech jobs, the companies hiring, and the people building the field.