Scalekit has put AgentKit on Product Hunt, and the pitch is narrow on purpose: an MCP gateway for production agents that act on your behalf. That phrase, “act on your behalf,” is doing the real work. It is not a claim about a smarter model or a faster runtime. It is a claim about authority. The product is less a model wrapper than a broker that sits between an agent and the systems it touches, and the fact that a company is now selling that broker as a discrete product tells you where the agent market has actually moved.
The Model Context Protocol, or MCP, is the connective tissue here. It is the standard that lets an agent call tools, read data, and reach external services through a defined interface. MCP has spread quickly because it solves an obvious problem: every lab and every tool vendor was inventing its own plug format, and nobody wanted to maintain N integrations for N agents. A shared protocol collapses that. But a shared protocol also creates a new problem, and this is the one AgentKit is aimed at. Once an agent can reach a dozen tools through one protocol, who decides which tools it may reach, under what conditions, and with whose credentials?
That is the gap between a demo and a deployment. A weekend agent that reads your calendar is a toy. An agent that files expense reports, opens pull requests, or moves money is an operating risk. The moment an agent “acts on your behalf,” it inherits your permissions, and your permissions are the thing every security team gets paid to protect. Scalekit’s framing, a gateway for production agents, is an admission that the hard part of agents is not intelligence. It is control.
What the gateway actually implies
A gateway is a chokepoint, and chokepoints are where the value accrues in infrastructure. The pattern is familiar from older layers of the stack. API gateways became the place where companies enforced rate limits, authentication, and logging, and the vendors that owned that chokepoint (Kong, Apigee, and the cloud-native incumbents) built durable businesses on it. An MCP gateway is the same idea one layer up. It routes an agent’s tool calls, and in routing them it can inspect, log, throttle, and deny them.
That is a defensible position, and it is also a contested one. The labs have every incentive to own the agent runtime themselves. Anthropic, which introduced MCP, ships its own tooling. OpenAI, Google, and Microsoft are all building agent frameworks with their own permission models. A standalone gateway has to justify why a customer would route sensitive tool calls through a third party rather than through the platform that already holds the model. The answer, if there is one, is neutrality: a gateway that works across Claude, GPT, Gemini, and whatever open-weight model a team runs locally, and enforces one policy across all of them. Whether buyers want that neutrality, or would rather take the path of least resistance inside one vendor’s ecosystem, is the open question AgentKit is betting on.
{/* TODO: verify Scalekit’s funding, founding date, and any named customers or pricing tiers; the Product Hunt listing does not supply these. */}
The economy of “acting on your behalf”
There is a quieter economic story here. Agents that act autonomously consume compute in bursts that are hard to forecast. A chat session is one request and one response. An agent that plans, calls tools, checks results, and retries can fan one user intent into dozens of model calls and external API hits. The gateway is where that fan-out becomes visible, and visibility is the precondition for billing, budgeting, and capacity planning. A company that can meter agent tool calls per team, per agent, per task, is a company that can sell the thing enterprises actually need before they scale agents: a cost ceiling.
This is why the “production” qualifier matters. Production agents fail in ways that demos do not. They loop. They call the wrong tool with the right arguments. They leak a credential into a log. They act on stale context and do something irreversible. None of these are model-quality problems in the narrow sense. They are orchestration and governance problems, and they get solved by infrastructure, not by a better prompt. Scalekit is selling into that gap, and the gap is real.
The risk is that the gap closes from both ends. On one side, the model providers keep absorbing orchestration into their platforms, and a thin gateway gets squeezed. On the other, open-source MCP proxies and self-hosted gateways will appear, because the protocol is open and the demand is obvious. A commercial gateway has to win on the things open source is bad at: audit trails that satisfy a compliance officer, SSO and role-based access that plug into an existing identity provider, and a support contract when an agent does something expensive at 2 a.m.
A gateway is a chokepoint, and chokepoints are where the value accrues in infrastructure. An MCP gateway is the same idea one layer up.
What this means for builders
If you are building agents, the lesson from a launch like AgentKit is not “buy this.” It is that the permission layer is now a first-class design decision, not an afterthought. The teams that ship reliable agents will be the ones that decided early which tools an agent may call, how those calls are authenticated, and how a human can see and stop them. You can build that yourself, and many teams will, at least at first. You can also buy it, and the pitch for buying it is the same pitch that sold API gateways a decade ago: the chokepoint is where the outages and the breaches happen, so put a real product there.
For the industry, the signal is that the agent conversation has quietly moved from capability to custody. The interesting question is no longer whether an agent can do a task. It is who holds the keys when it does. Scalekit is one of a growing number of companies answering “we do,” and the market will decide how many custodians the stack can support. The thing to watch is not the Product Hunt upvotes. It is whether any of these gateways lands a named enterprise customer with a real compliance requirement, because that is the moment the category stops being a pitch and starts being a line item.