Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Anthropic Just Bought a Dev Tools Startup for $300M. Here's What Its Founder Told Me.
TL;DR
MCP is promising, but today it breaks down on context and usability — Alex Rattray says the dream is letting an AI do anything a human could do across tools like Stripe, Salesforce, Slack, and Gmail, but exposing hundreds of API endpoints blows the model’s context window and confuses it.
Good MCP servers are hand-crafted products, not API wrappers — The teams succeeding today keep tool counts low, use extremely precise names and descriptions, minimize input parameters, and trim responses down to only what the model needs, rather than dumping a raw OpenAPI spec into MCP.
Stainless thinks code execution beats giant tool menus — Rattray’s bigger bet is giving models two tools instead: one to execute TypeScript against a typed SDK and one to search docs, which avoids massive context costs and lets the model iterate with type-checking instead of hallucinating raw HTTP calls.
Security has to live at the API layer, not just in MCP packaging — He argues the real control point is OAuth with granular scopes on the API itself, because hiding capabilities in an MCP server doesn’t matter much if the underlying API can still do dangerous things.
Rattray already uses MCP more for operating Stainless than for coding — He queries a read-only Postgres copy plus Notion, HubSpot, and Gong to ask things like “what interesting customers signed up last week?”, then has Claude Code save customer quotes and refined SQL queries into an internal git-based knowledge repo.
The long-term thesis is ‘cyborg’ software: LLMs plus traditional code — Instead of pure chat or pure automation, Rattray sees the future as AI systems that explore with natural language, execute real code when needed, and then promote repeated one-off actions into durable software and dashboards.
The Breakdown
From barefoot runs in Philly to the API layer of the internet
Dan opens with a warm friend-intro: Alex Rattray, founder and CEO of Stainless, ex-Stripe, college co-conspirator, and famously a guy who used to run barefoot through Philadelphia because shoes kept wearing out faster than his feet. It’s goofy, but it sets the tone: this is a real conversation between two builders who’ve watched each other grow from startup kids into infrastructure people shaping how AI talks to software.
APIs as the internet’s dendrites — and why AI changes the equation
Rattray explains APIs as the way computers talk to computers, calling them the “dendrites of the internet.” Stainless started with SDKs for companies like OpenAI and Anthropic, but now the new question is: if Python developers get ergonomic SDKs, what’s the equivalent interface for LLMs? MCP is the current answer, but he says it still falls short of the big vision.
The real MCP dream: let AI do the annoying 15-click workflow
His example is intentionally mundane: refund Dan for stripey socks, generate a discount code, and send a note. A human does that by hopping across five apps and 15 clicks; the promise of agentic AI is typing one request into Claude or ChatGPT and having it just happen. The catch is that to do this generally, the AI would need access to basically everything a human can do in each SaaS dashboard — which means hundreds of tools, endpoints, and docs.
Why MCP gets hard fast: context windows, black boxes, and awkward tool design
The problem isn’t just scale, it’s ergonomics. If you expose every Stripe endpoint as a tool, you’ve already burned huge amounts of context before the model even starts reasoning, and pagination or bulky responses make it worse. That’s why Rattray says good MCP today is more like product management plus evals plus careful writing: fewer tools, sharper descriptions, fewer parameters, and tiny responses, sometimes with tricks like JQ filters.
A practical hack: dynamic mode and real-world internal use
Stainless has a “dynamic mode” with just three tools — list endpoints, inspect an endpoint, execute an endpoint — which scales better on context but adds latency because the model needs multiple turns just to do one task. Rattray says he personally uses MCP more on the business side than for coding, combining Notion, HubSpot, Gong, and a read-only Postgres database to ask who the interesting new Stainless customers were last week and get a synthesized answer across systems.
Claude Code as a company memory system
One of the most useful habits he’s developed is asking Claude Code to write things back into a git repo: customer quotes with citations, useful SQL queries, and other internal research artifacts. It’s not polished, and he explicitly says the structure is messy, but that’s kind of the point — unstructured markdown is already useful enough if the model can search it later. He used this loop during board prep, refining SQL queries with business context and then saving the best version for next time.
The bigger bet: replace giant tool catalogs with code execution
Rattray’s main thesis is that the future of AI is “cyborgs” — part LLM, part traditional software. Instead of handing the model 500 tools, give it a code execution environment plus doc search, let it write TypeScript against a typed SDK, and have that code run next to the API in the cloud. That keeps upfront context tiny, avoids sending paginated haystacks back into the model, and uses static typing to catch mistakes like hallucinated methods before they hit production.
Security, YOLO adoption, and what Stainless is building next
On security, he’s clear: the right place for control is the API layer, with OAuth and granular scopes, not just restricting MCP exposure. Dan pushes him on whether Stainless should ship a more permissive, developer-first version quickly — invoking Stable Diffusion and Claude Code’s “dangerously skip permissions” as examples of YOLO products winning early adoption — and Rattray basically says yes, they’re working on it, starting with secure single-provider sandboxes and expanding from there.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
