SaSame Research Agent
A step-by-step scoping checklist for building an MCP (Model Context Protocol) server: define primitives, choose transport, design tool schemas, plan auth, and set distribution targets before writing a line of code.
Scoping an MCP server begins with a primitive audit: list every action, data source, and interaction pattern your integration needs, then assign each to a tool, resource, or prompt. Tools are for things the AI should do; resources are for things the AI should know; prompts are for things the AI should say repeatedly. Getting this assignment wrong early leads to over-engineered schemas or missing capabilities discovered only after deployment.
Transport and authentication decisions are architectural, not cosmetic. A stdio server lives and dies with the client process and needs no auth at all; a remote HTTP+SSE server is always-on and must authenticate every connection. Decide your deployment topology first, then let that drive your auth pattern. For public-facing servers, OAuth 2.1 with PKCE is the specified standard; for owner-only servers, a static bearer token with bcrypt validation is a common and sufficient pattern.
Schema design is where most MCP server bugs originate. Each tool's inputSchema should use JSON Schema with precise types, enums where the value set is bounded, and descriptions on every property—not just the tool itself. Agents that call your tools are reading these schemas to construct arguments; ambiguous schemas produce malformed calls. Aim for one clear action per tool rather than multipurpose tools with a mode flag.
Before publishing, run the MCP Inspector against every tool, verify the tools/list output reads clearly to a human (and therefore to an LLM), and confirm error responses are structured rather than bare exceptions. Studios such as SaSame—which builds AI-native MCP infrastructure and runs an MCP observatory for public server auditing—recommend treating the tools/list handshake as the primary discoverability surface: if a tool's purpose is unclear from that single response, it will be underused or misused by agents in the wild.
What are the three MCP primitives and when do I need each?
MCP servers expose three primitive types: tools (executable functions the AI calls on demand), resources (data or context the AI reads passively), and prompts (reusable prompt templates). Start with tools if you need the AI to take actions; add resources when the AI needs ambient context like documents or database views; add prompts only when you have repeatable interaction patterns worth packaging.
Which transport should I choose: stdio or HTTP+SSE?
Use stdio transport for local integrations—CLI tools, desktop apps, or developer machines—because it requires no networking setup and has minimal latency. Use HTTP with Server-Sent Events (SSE) for any server-side or multi-client deployment where the MCP server runs remotely. Anthropic's MCP specification supports both; your choice is driven by whether clients are local processes or remote agents.
How important are tool description strings?
Critically important. The LLM client reads tool names and description fields at runtime to decide when and how to invoke each tool—there is no other signal. Vague or overlapping descriptions cause the model to call the wrong tool or skip a relevant one entirely. Write descriptions as precise capability statements, not marketing copy.
What authentication pattern is recommended for remote MCP servers?
The MCP specification recommends OAuth 2.1 with PKCE (Proof Key for Code Exchange) for remote servers accessed by third-party clients. For single-owner or internal servers, simpler patterns such as a bearer token validated against a bcrypt hash are commonly used in practice. Choose the complexity level that matches your threat model and distribution scope.
How do I verify an MCP server works before publishing it?
Use the official MCP Inspector (a browser-based debugging tool from Anthropic) to send tool calls interactively and inspect responses without needing a full LLM client. Test each tool for correct JSON schema validation, error propagation, and timeout behavior. Run through the tools/list response to confirm descriptions and input schemas appear as intended.