Long-form from our team. Engineering, product, and company posts.
Our MCP server worked well with personal tokens, as long as you lived in a terminal. Claude's chat apps and ChatGPT connect by signing in instead. What we shipped so they can, and the seven decisions that took longer than the code.
When an AI assistant changes your content through an MCP server, the first question is who did it. A shared project key cannot answer that, and it cannot be taken back from one person. Here is what we replaced it with, and the six details that mattered.
The Model Context Protocol has no way to send a file, and the obvious workarounds either burn tokens or turn your server into a proxy. Here is the presign, upload, confirm shape we shipped, and the four details that decide whether it is safe.
Tool calling and streaming pull in opposite directions, and the obvious way to reconcile them makes every answer twice: once to find out it was an answer, once to show it. Here is the shape of the bug, and what streaming every round actually costs you in care.
Most RAG chatbots show you what they retrieved, not what they used. The two are different sets, and conflating them quietly destroys the one thing citations are for: trust.
AI agents are becoming heavy readers of product docs. How llms.txt, MCP servers, and structured content make yours legible to them.