Make Products AI-Usable, Not AI-Embedded

haltakov · x · 2026-07-16

The author argues that many SaaS products are ruining both costs and user experience by forcing AI copilots into their interfaces: users already pay for ChatGPT/Claude, yet are asked to buy another layer of expensive, restricted AI within each product. They propose a better approach: **make your product accessible to users' existing AI assistants** rather than embedding your own. This can be done by exposing data, tools, workflows, and instructions to a **remote MCP server**, allowing users to connect via their preferred ChatGPT, Claude, or other agents. Key takeaways include: - Remote MCP can use standard browser auth flows: simply connect, log in, and authorize. - It can provide instructions and interactive UIs, making the experience feel like a real app rather than a developer tool. - It makes economic sense: the product provides domain data, permissions, storage, and knowledge, while users bring their own frontier models, preventing SaaS providers from buying tokens only to repackage them as "AI credits" for resale. - While not applicable to every scenario, the author believes developers are asking the wrong question: instead of "how to add AI to our product," they should ask "how to make our product usable by users' existing AI."

Original post →

More from coding & agent

coding & agent channel →