MCP gives models a standard way to connect to services, but the ecosystem still has major problems for users and developers.
2
Finding, installing, securing, hosting, distributing, observing, and charging for MCPs are all unresolved parts of the system.
3
Smithery is building an AI gateway and orchestration layer, demonstrated through an agent that finds a GitHub issue and creates a Linear ticket.
Summary
Henry Mao argues that better model intelligence has not produced useful autonomous agents because intelligence remains disconnected from context and action. He sees MCP as a promising open standard for connecting models to services, but says the ecosystem has quickly developed problems around fragmentation, installation friction, security, payments, hosting, tooling, distribution, observability, and monetization. Smithery is his response: an AI gateway intended to help agents find and use curated services. In the demo, an agent searches a GitHub repository for its most pressing issue, then creates a detailed ticket in Linear through separate MCP connections. Mao is clear that this only covers part of the problem. He expects the internet to involve more tool calls than clicks, with the quality of an agent's experience becoming more important than the traditional user experience.
Model intelligence is still trapped without context and capability
Mao begins with the gap between benchmark performance and practical usefulness. Humans perform at about 80% on the ARC AGI challenge, while language models struggled with it for years. OpenAI's o3 then reached human-level performance on ARC AGI 1 and did well on math benchmarks. That progress made Mao ask where the autonomous agents were that could handle unwanted tasks. He describes this as Claude's paradox: frontier labs can create powerful intelligence, but the intelligence is stuck in a box. Agents need useful inputs and outputs, which means access to context and capabilities.
MCP standardizes connections without solving the whole ecosystem
Anthropic released the Model Context Protocol in November 2024 as an open standard for connecting language models to different services. Mao was excited because MCP could standardize how models talk to services, and a small developer community was already building around it. He quickly rejects the idea that this finishes the work. Once services begin targeting agents, new user and developer problems appear around discovery, installation, security, hosting, tool design, distribution, observability, and revenue. MCP creates a common connection method, but the surrounding systems still need to be built.
Users need better discovery, installation, security, and payments
Mao lists several problems for MCP users. The number of servers is growing, which makes it difficult to find high-quality ones. An official registry is being developed, but a registry alone does not explain how servers should earn reputation. Installation often requires a five-step process from a GitHub repository, and users may unknowingly install an insecure server. The economy for AI-native services is also unsettled. Mao asks how agents can pay on a user's behalf and how users can avoid subscribing to a hundred separate services that each charge $10 per month.
MCP developers still lack the infrastructure and feedback they need
Developers face their own set of problems. Hosting MCPs remains difficult even though Streamable HTTP has made it easier to find hosting platforms. Developers still have to handle stateful sessions, resumability, and related concerns. The official MCP Inspector provides basic testing for tools and prompts, but Mao says builders still need answers about how to design an effective MCP, predict whether an agent will call a tool, and create a good agent experience. Distribution is unresolved because developers need a way for agents to discover their MCPs. Observability is also missing: after launch, developers need to understand usage so they can improve the service and decide how to make money from it.
Smithery is building an AI gateway around these missing pieces
Mao says Smithery started in December 2024 to address the problems on both sides of the ecosystem. He describes it as an AI gateway intended to grow and orchestrate an era of AI-native services for agents. He does not claim the company has solved the whole space. Instead, Smithery has built a playground to show what becomes possible when an agent can access thousands of curated MCPs. The playground focuses on giving an agent a way to find the right services and combine them in an end-to-end task, rather than asking a user to configure every connection manually.
The Smithery demo combines GitHub and Linear through an agent
The demo starts with a request to find the most pressing issue in the Smithery CLI GitHub repository and create a new ticket in Linear. The agent first searches available servers, connects to the best match, uses the GitHub MCP to inspect bugs, and identifies the highest-priority item. It then creates a Linear ticket with detailed information and a link back to the original GitHub issue. Mao presents this as one task solved across two MCPs. He also says the example only scratches the surface of what could happen as more servers become available through Smithery.
Mao expects tool calls and agent experience to shape the internet
Mao closes by pointing to the growth in deployed servers and tool calls shown in a graph. He says developer enthusiasm gives him confidence that the ecosystem can address the problems he listed. His broader prediction is that tool calls will dominate the future of the internet more than clicks do. In that setting, the experience of the agent matters more than the traditional user experience. He says this experience will be built by the wider developer community, rather than by Smithery or a small group of companies.
"How do I get agents to pay on my behalf? And how do I avoid subscribing to 100 different services each charging $10 a month?"03:21
Who should watch
You are building or evaluating MCP servers and need a clear list of the ecosystem problems beyond the protocol itself.
Your team wants agents to discover and combine external services without asking users to manage every installation and connection.
You are thinking about hosting, observability, distribution, or payments for AI-native services.
