Inside Stainless: The Developer Tools Startup Anthropic Just Bought for $300 Million (51 min)
ai-driven-innovation-economy
existential-ai-risks
international-ai-agreements
- Release date: 2026-05-20
- Listen on Spotify: Open episode
- Episode description:
If your MCP server has dozens of tools, it's probably built wrong. You need tools that are specific and clear for each use case—but you also can't have too many. This creates an almost impossible tradeoff that most companies don't know how to solve.That's why we interviewed Alex Rattray, the founder and CEO of Stainless. Stainless builds APIs, SDKs, and MCP servers for companies like OpenAI and Anthropic. Alex has spent years mastering how to make software talk to software, and he came on the show to share what he knows. We get into MCP and the future of the AI-native internet. [Disclosure: Dan is a small investor in Stainless.]If you found this episode interesting, please like, subscribe, comment, and share.To hear more from Dan Shipper:Subscribe to Every: https://every.to/subscribeFollow him on X: https://twitter.com/danshipperGet started with Braintrust at https://www.braintrust.dev/ Timestamps: 00:01:15 - Introduction 00:05:09 - APIs and MCP, the connectors of the new internet 00:11:00 - Why MCP exists 00:17:15 - Why MCP servers are hard to get right 00:20:24 - Design principles for reliable MCP servers 00:25:06 - Using MCP for business ops at Stainless 00:40:57 - Alex's take on the security model for MCP 00:44:42 - How one-off AI actions become permanent production softwareLinks to resources mentioned in the episode:Alex Rattray: Alex Rattray (@RattrayAlex), Alex RattrayStainless: https://www.stainless.com/
Summary
- ⚙️ MCP Scaling Problem: Exposing full APIs as MCP tools consumes the entire LLM context window, making it unusable for complex tasks.
- 💻 Code Execution as Solution: A code execution tool (write code, run it on the API provider’s server) uses far less context, runs faster, and is more natural for LLMs than hundreds of tool definitions.
- 🔒 Security at API Layer: Proper MCP security requires OAuth with granular scopes at the API layer, not simply limiting exposed tools.
- 🔍 Practical MCP Uses Today: MCP is already used for business queries across Notion, HubSpot, and databases, and for building a git-based knowledge repository for AI.
- 🚀 YOLO Adoption Strategy: Startups should prioritize speed and risk-taking in AI adoption (like ‘YOLO mode’) over excessive caution to gain early traction.
Insights
- Why is scaling MCP to full API complexity so difficult?
- Time: 10:58 – 14:48
- Answer: The vision of MCP is to let an AI do any task a human operator can, across many apps, by exposing every API endpoint as a tool. But the Stripe API alone has hundreds of endpoints, and translating that into MCP tools consumes the entire context window of an LLM in just tool definitions, leaving no room for the actual task or conversation. So the naive approach fails because models cannot hold that much context or navigate that many options effectively.
- How can MCP servers be designed well today?
- Time: 20:24 – 23:10
- Answer: Good MCP design today requires keeping the number of tools small and names/descriptions very precise, using few input parameters, returning minimal response data, and building a feedback loop to the server. It functions like product management for AI: understanding user needs, engineering good ‘handcrafted’ wrappers, and running evals across different clients and models. This is harder than designing for human developers because we can’t ‘think like an LLM’ easily.
- What practical use cases for MCP are already working?
- Why does code execution outperform massive tool sets for AI agents?
- Why is security in MCP best handled at the API layer, not the MCP layer?
- How do startups best adopt AI: cautious or ‘YOLO’?