MCP servers

Build Smarter MCP Servers by Keeping Orchestration in the Host

Build smarter MCP servers by keeping orchestration in the host, not the tool.

3 min readTowards Data Science

The clearest line of progress in AI-native systems is the one that draws a boundary. A recent demonstration of a Python SDK v2 catalogue server, with two narrow tools, one resource, and one prompt, makes the case that orchestration belongs in the host, not scattered across every server. This is the right call, and it matters more than most architectural debates because it directly affects how teams build tools they can actually trust and maintain. The post, published on Towards Data Science, shows a server designed with deliberate restraint: it handles only what it must and leaves planning, workflow state, and authorization to the layer above. That restraint is the feature.

For anyone building agentic systems, the temptation is to let each server grow into its own mini-orchestrator. A server that manages its own workflow state seems convenient until you try to debug a chain of five such servers, each holding its own idea of what happened and why. The catalogue server example sidesteps that entirely. It offers two tools, one resource, and one prompt, nothing more. The host decides the order of operations, tracks progress, and enforces permissions. This aligns directly with the principles outlined in our recent piece on Explore agentic system design with a practical 7-step roadmap, where clear separation of concerns between tools, memory, and control loops emerged as a core pattern. A server that tries to do everything becomes the system's bottleneck and its single point of failure.

The practical consequence for data teams is simpler maintenance and clearer failure modes. When a tool misbehaves, you know exactly where the problem lives. When authorization needs updating, you change it in one place. This also makes the server more reusable across different hosts and use cases, because it carries no assumptions about the workflow it serves. We saw a similar lesson in Make Your Analytics Work Irreplaceable by Building Tools Teams Trust, where following data from input to output revealed that the most trusted tools were the ones with the narrowest, most predictable behavior. A catalogue server that returns products by category and nothing else is boring by design, and that is exactly what makes it reliable.

What remains an open question is how far this principle scales. The demonstration server is deliberately small, but real-world systems will need servers that handle more complexity while still leaving orchestration to the host. The boundary between a tool and a mini-orchestrator can blur when a server needs to coordinate internal state across multiple calls. The specific detail to watch is how future SDKs handle that pressure, whether they enforce the separation or make it easy to violate. For now, the rule is simple: if your server is deciding what to do next instead of just doing what it is told, you have moved orchestration to the wrong place.

From Towards Data Science

A Python SDK v2 catalogue server with two narrow tools, one resource, and one prompt shows how to keep planning, workflow state, and authorisation host-side.

The post How to Design MCP Servers That Leave Orchestration to the Host appeared first on Towards Data Science.

Read the original at Towards Data Science