Building MCP Servers for ChatGPT: Best Practices on Tools, Schemas, Annotations, and Server Instructions

Thomas Alandi from OpenAI's Plugins team presents practical guidance for developers building MCP servers for ChatGPT, drawing on questions from the developer community and usage data from OpenAI's public directory. The talk addresses how to structure tools, schemas, annotations, and server instructions to improve discoverability and conversation quality.

■ Tools
- How many tools to expose to the model and why there is no hard limit
- Common anti-patterns such as hiding real tools behind a generic command list

■ Schemas and Annotations
- Keeping schema sizes reasonable and reviewing them for errors
- Why annotations are required for submission and how they affect auto-approval vs. destructive-tool confirmation

■ Server Instructions and OpenAI's MCP Contributions
- Recommended length and placement of server instructions
- OpenAI's MCP apps, ChatGPT Add-Ons, event support, and file upload unification

Geared toward developers publishing MCP servers or ChatGPT plugins, the talk offers actionable best practices and insight into how OpenAI evaluates submissions. Viewers can apply these guidelines to improve their tool design and increase the chances of successful directory integration.