Inside Stainless: The Developer Tools Startup Anthropic Just Bought for $300 Million

Overview

This episode dives into the future of AI-native internet infrastructure, focusing on MCP (Model Context Protocol) servers and their role in enabling AI systems to interact with APIs and software. Alex Rattray, CEO of Stainless, shares insights on designing effective MCP servers, overcoming their challenges, and the broader implications for AI-driven automation and software development.

Notable Quotes

- The future of AI is cyborgs—part neural net, part traditional software. – Alex Rattray, on the hybrid nature of AI systems.

- The longer you wear your shoes, the more worn out they get, but the longer you wear your feet, the tougher they get. – Alex Rattray, humorously reflecting on his barefoot running days.

- We haven’t figured out how to expose an API ergonomically to an LLM the way we’ve cracked it for Python developers. – Alex Rattray, on the challenges of designing AI-friendly APIs.

🧠 Understanding MCP and Its Challenges

- MCP (Model Context Protocol) allows AI models to interact with APIs as if they were users, enabling automation of tasks like sending emails or processing refunds.

- Alex Rattray explains that while MCP is promising, it struggles with scalability, context window limits, and security concerns.

- Current MCP implementations often require handcrafted tools tailored to specific use cases, which limits flexibility and increases complexity.

- A major challenge is balancing the number of tools: too many overwhelms the model, while too few limits functionality.

⚙️ Principles for Designing Effective MCP Servers

- Keep the number of tools small and descriptions precise, even though these goals are inherently at odds.

- Limit input parameters and response data to only what the model needs, reducing cognitive load and context usage.

- Use techniques like JQ filters to streamline JSON responses and improve efficiency.

- Stainless employs a dynamic mode with three core tools—list endpoints, get endpoint details, and execute endpoint—to simplify interactions while maintaining flexibility.

🚀 AI in Business Operations

- Alex Rattray shares how Stainless uses MCP servers internally for tasks like analyzing customer data across Notion, HubSpot, and Gong.

- He describes a workflow where AI generates SQL queries, iterates based on feedback, and stores them for future use, streamlining board prep and analytics.

- Stainless also experiments with AI-driven bug fixes, where models attempt to resolve customer issues autonomously, though success rates are still evolving.

🔒 Security and the Future of AI Tooling

- Security in MCP should be enforced at the API layer using OAuth with granular permissions, rather than limiting MCP tool exposure.

- Stainless is exploring secure code execution environments where AI can write and execute TypeScript code directly, interacting with APIs without exposing sensitive data.

- This approach minimizes context window usage and allows for more complex, multi-step operations without overwhelming the model.

🛠️ From One-Off Actions to Production Software

- AI's ability to execute code dynamically could transform how software is developed, with one-off tasks evolving into reusable production tools.

- Alex Rattray envisions a future where AI-generated code is committed to repositories, automating repetitive tasks and enabling seamless transitions from exploration to production.

- This shift could reduce the need for traditional tool development, with prompt engineering becoming the primary focus for enabling AI functionality.