--- schema: 1 type: topic title: MCP and AI Infrastructure description: A synthesis of the channel's material about MCP and AI Infrastructure. canonical: >- https://liquidgold.cc/channel/everyinc/explore/topic/mcp-and-ai-infrastructure-frastructure markdown: >- https://liquidgold.cc/channel/everyinc/explore/topic/mcp-and-ai-infrastructure-frastructure.md llms: >- https://liquidgold.cc/channel/everyinc/explore/topic/mcp-and-ai-infrastructure-frastructure/llms.txt parent: 'https://liquidgold.cc/channel/everyinc.md' channel: everyinc updated: '2026-10-23T22:12:35.346Z' language: null children: [] topicSlug: mcp-and-ai-infrastructure-frastructure searchChunks: [] --- # MCP and AI Infrastructure A synthesis of the channel's material about MCP and AI Infrastructure. Canonical: https://liquidgold.cc/channel/everyinc/explore/topic/mcp-and-ai-infrastructure-frastructure Parent: [Every](https://liquidgold.cc/channel/everyinc.md) After this chapter you will be able to connect agents to external tools and data using MCP servers, explain the infrastructure the agentic web is being built on, and evaluate the tradeoffs — context cost, security, reliability — that separate toy demos from production agents. The material synthesised here comes from interviews with the people actually building this stack: Alex Rattray of Stainless, Microsoft CTO Kevin Scott, Anthropic's Claude platform leads, Salesforce's Silvio Savarese, and Stripe's Emily Glassberg Sands. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@334] [^385ed583-4107-5c5a-b47a-33e1dde27a3b@327] [^90e39a1e-87e8-5b3f-b9f2-f003e83fd62b@96] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@220] [^4872db36-e98a-5f6b-bc5c-b11e48c9411b@139] [^ed15d1a6-8e60-53ca-9dc0-bd0a82a2ddf5@113] [^eea0a522-b6b8-5472-bf62-08b0a42ab8a3@268] [^ea9761fa-0184-555e-be30-4928a6682680@327] ## Agents need infrastructure to reach the internet An **API** is how one computer program talks to another. Alex Rattray's analogy is worth internalising: APIs are the _dendrites_ of the internet — the connection points where programs, like neurons, exchange information. Without those connections there is no thought, and no internet. > APIs are the dendrites of the internet. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@334] [^ea9761fa-0184-555e-be30-4928a6682680@327] For agents to be useful, Kevin Scott argues, they must **take action on your behalf** — using tools, changing systems, consulting rich information sources — which requires an ecosystem that looks like the internet, where existing websites and APIs get plumbed through so agents can talk to them. [^eea0a522-b6b8-5472-bf62-08b0a42ab8a3@268] ### MCP as the HTTP of the agentic web Scott's central analogy: **MCP** plays the role in the agentic web that **HTTP** plays on the internet, and Anthropic's **NLWeb** aims to be the moral equivalent of **HTML** — simple, composable, open protocols that layer together and reach ubiquity. [^eea0a522-b6b8-5472-bf62-08b0a42ab8a3@326] [^385ed583-4107-5c5a-b47a-33e1dde27a3b@327] Standard protocols also fight **Conway's law** — you ship your org chart. Scott is pushing for every internal Microsoft system to speak a standard agent protocol so team boundaries don't leak into agent architecture. [^eea0a522-b6b8-5472-bf62-08b0a42ab8a3@423] ### Why MCP matters: an SDK for LLMs Rattray frames MCP by analogy to SDKs: SDKs give developers a convenient way to interface with an API — think `pip install stripe` — so what gives **LLMs** an easy way to interface with an API? MCP is the obvious answer, but as it rolls out, it isn't working so great. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@536] The stakes are visible in the _stripey-socks refund_ example: one human request — refund Dan for the socks he bought yesterday — normally spans five apps and fifteen screens. The agentic vision is typing that sentence into ChatGPT or Claude and having it done natively. [^ea9761fa-0184-555e-be30-4928a6682680@780] ## The context-window problem and MCP design Translating an API one-to-one into MCP tools fails because big APIs are big. The Stripe API exposes **hundreds of endpoints**; exposing each as an MCP tool — with descriptions, request and response properties, and docs — burns hundreds of thousands of tokens from the context budget _before any real work starts_. It's a poor use of context and simply confusing. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@871] [^ea9761fa-0184-555e-be30-4928a6682680@919] ### Concrete design rules Building an MCP server well takes real **product management** — talk to customers, watch how they work — plus serious engineering and an **evals system** across the matrix of clients (Cursor, Claude Code, etc.) and underlying models. Feedback from the client back to the server is still a black box; even a _send-feedback_ tool is on the table. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@1220] Rattray's recurring rules: - keep the number of tools **small** - make tool names and descriptions **precise and specific** — directly at odds with rule one, because good writing is hard - keep few, concisely-described **parameters** in the input schema - return only what the model needs, using a **JQ filter** to trim JSON responses [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@1346] ### Dynamic mode: a scaling tradeoff To solve tool-browsing at scale, Stainless offers **dynamic mode**: no matter how big the API, the MCP server exposes just three tools — _list endpoints_, _get endpoint_, _execute endpoint_. Context cost scales beautifully, but one action costs three turns of the model — slower, more expensive, and slightly lossier in performance. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@1453] ## The code-execution future and security at the API layer The alternative to a bajillion tools is **code execution**: the model gets two tools — one to run TypeScript against the API's SDK, one to search the docs. Finding Dan among paginated transactions becomes nested for loops that run server-side, logging only the found customer and created refund — about ten lines of context, ~1,000 tokens up front, fast, with no round trip to the model per step. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@2142] [^ea9761fa-0184-555e-be30-4928a6682680@2146] ### Static typing as the hallucination guard The catch: code execution only works if the library works perfectly out of the box — right version, no hallucinated methods. **Static typing** is the safety net: a type checker catches errors in real time — `stripe.transactions.list` doesn't exist; did you mean payment intents, orders, or balance transactions? — returning relevant documentation inline, powered by a sub-agent unburdened by the full conversation context. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@2398] [^ea9761fa-0184-555e-be30-4928a6682680@2412] ### Security: the API layer, not the MCP layer Rattray is direct on a contested point: limiting what MCP exposes is the wrong lever, because anything in the underlying API remains reachable under the hood. Security has to take place at the **API layer itself** — OAuth with granular permissions and proper scopes. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@2454] For sandboxed code execution, Stainless starts with **one API provider per sandbox**, locking network access so nothing can connect out except the target — `api.stripe.com` and nothing else. [^913c216d-5ccb-51d9-8f42-8feb7c017d9d@2611] [^ea9761fa-0184-555e-be30-4928a6682680@2638] Scott sketches a complementary, community-level direction — still uncertain even in his own telling: **agent identities**, entitlement systems, and permission requests to both delegating users and administrators, built openly on top of MCP rather than proprietary to Microsoft. [^385ed583-4107-5c5a-b47a-33e1dde27a3b@639] [^eea0a522-b6b8-5472-bf62-08b0a42ab8a3@632] Both also push back on the framing that open ecosystems must sacrifice robust security — Scott calls open-versus-verticalized possibly a **false dichotomy**. [^385ed583-4107-5c5a-b47a-33e1dde27a3b@763] ## Production infrastructure, orchestration, and the agent economy Teams expect **harness engineering** to be the bottleneck. Katelyn Lesse says the real wall everyone hits is **production infrastructure** — keeping servers running, storing transcript data, and secure sandboxing, where a dying sandbox takes your whole agent with it. Harness choice matters too: harnesses performed _drastically differently_ on evals, and new models require harness and model to be _very paired_ rather than hot-swappable. [^4872db36-e98a-5f6b-bc5c-b11e48c9411b@1181] [^4872db36-e98a-5f6b-bc5c-b11e48c9411b@813] [^4872db36-e98a-5f6b-bc5c-b11e48c9411b@692] ### Managed infrastructure and why it matters This is why **managed agent platforms** may be sneaky important: Anthropic ships a scalable cloud-hosted agent platform with memory and orchestration, while — as the hosts argue — OpenAI has only an Agents SDK with no cloud version yet. Inside a bigger company, a **Mac mini going down** is unacceptable; **reliability**, not peak performance, is the real value. [^ed15d1a6-8e60-53ca-9dc0-bd0a82a2ddf5@211] [^ed15d1a6-8e60-53ca-9dc0-bd0a82a2ddf5@342] ### When agents fail for business reasons Silvio Savarese decomposes an agent into four parts — **memory** (typically RAG), a **brain** or reasoning engine, **actuators** (API and function calls), and an **interface** — and argues an agent without connected data and context is useless. Enterprise deployments stumble on **siloed data**: sales and marketing manage data differently, and AI cannot operate well when the dots aren't connected. He also names **jagged intelligence** — models that prove theorems yet fail at common-sense business tasks — and uses simulated organizations with synthetic data to stress-test agents, with the reality gap still open work. [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@220] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@342] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@431] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@485] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@676] ### The economics underneath: agents as new economic actors Emily Glassberg Sands frames the macro shift: the internet was built on the assumption that a **person** is the main actor. Agents are now a new class of internet actor, forcing every layer of the stack — discovery, payments, identity, fraud — to be rethought. [^90e39a1e-87e8-5b3f-b9f2-f003e83fd62b@96] - **Fraud targets compute**: fraudsters steal free-trial credits rather than card numbers. One large AI company saw only 4% of free trials convert, each trial costing $25 in LLM spend — $625 per payer — with most trial users abusers. [^90e39a1e-87e8-5b3f-b9f2-f003e83fd62b@240] [^90e39a1e-87e8-5b3f-b9f2-f003e83fd62b@374] - **Defenses move up-funnel**: fraud is now a full-funnel, customer-lifecycle problem, integrated at sign-up, not just at checkout. [^90e39a1e-87e8-5b3f-b9f2-f003e83fd62b@619] - **Infrastructure for agentic commerce**: the Agent Commerce Protocol (co-created with OpenAI), **shared payment tokens** that pass credentials merchant-ward without the agent ever seeing them, and Stripe's Link wallet letting consumers decide which agents may request credentials and under what limits. [^90e39a1e-87e8-5b3f-b9f2-f003e83fd62b@2760] [^90e39a1e-87e8-5b3f-b9f2-f003e83fd62b@3139] ### Interaction is going asynchronous Across all of these sources, the direction of travel agrees: interaction shifts from **synchronous** prompting to **asynchronous** delegation. You hand a task off; the agent spends significant time making calls, integrating responses, and iterating, then reports back. Scott predicts it, Savarese predicts agents talking to agents needing defined protocols and guardrails, and Anthropic's Managed Agents launched with outcome-based tasks where the agent works asynchronously and returns only when done. [^eea0a522-b6b8-5472-bf62-08b0a42ab8a3@1551] [^50a3b3fa-9ad2-59a3-9c76-ddfe74635066@754] [^ed15d1a6-8e60-53ca-9dc0-bd0a82a2ddf5@113] - MCP is the standard protocol layer connecting agents to tools and APIs — Rattray calls it an SDK for LLMs; Scott calls it the HTTP of the agentic web. - Naive MCP design fails at scale: translating hundreds of API endpoints into tools burns hundreds of thousands of tokens of context before work begins; keep tools few, descriptions precise, schemas small, and consider dynamic mode (list/get/execute) with its three-turn tradeoff. - The emerging alternative is code execution against typed SDKs in network-locked sandboxes, with type checkers catching hallucinated calls — and security enforced at the API layer via OAuth scopes, not at the MCP layer. - The hard wall in production is infrastructure — uptime, transcript storage, sandboxing, reliability — not harness engineering; this is why managed, cloud-hosted agent platforms matter. - Agents are becoming new economic actors, which is forcing fraud defense (compute theft), pricing (tokens vs outcomes), and payment infrastructure (shared payment tokens, delegated purchasing) to be rebuilt. Connecting tools, data, and agents through protocols, sandboxes, and reliable platforms is what turns the agentic promise into infrastructure you can actually build on. \ ## Source videos > Missing: no public cited source videos are available.