You can save money on LLM tokens as a developer with MCP / ChatGPT apps (www.mikeborozdin.com)

Developer insight: packaging an app as an MCP (Model-Connected Plugin) or a ChatGPT app can cut your LLM token bill and open a distribution channel via the ChatGPT app store. Instead of your front-end or backend separately calling a model to synthesize prompts or structured data, an MCP runs “inside” the host LLM: the platform’s model generates the structured input and invokes your tool directly. That means you don’t need to pay for an extra model call from your server to parse user intent or assemble content, so token costs for those steps disappear. Key technical detail: MCPs require an explicit input schema and tool description so the host LLM knows what to provide. The blog shows a Zod-style schema for a language-learning audio-dialogue generator (array of speaker objects with personType, personName, text). In practice that shifts work from making raw prompt calls to designing precise schemas and handlers. Benefits include lower token spend and easy discoverability via the app store; trade-offs include relying on the platform’s execution, governance and rate limits, and designing robust schemas and validation for safety and correctness.