A few months ago I drafted a post called “From API-First to MCP-First,” about Amazon layering the Model Context Protocol on top of the Bezos mandate. I never published it. What AWS shipped since is more instructive than the slogan, and it points the other way.
The draft I shelved
In 2002 Jeff Bezos told every team at Amazon to “expose their data and functionality through service interfaces,” designed “from the ground up to be externalizable,” or be fired (Yegge’s retelling). That memo became AWS. So when Amazon started shipping MCP servers in 2025, the story wrote itself: API-first begets MCP-first, one tool per operation, “just add a server.”
I am glad I sat on it. Here is what actually happened.
What Amazon shipped
July 2025: an open-source AWS API MCP Server that turns natural language into “syntactically correct and valid CLI commands.” Not fifteen thousand tools. One tool that takes a command.
December 2025: Anthropic donates MCP to the new Agentic AI Foundation, and AWS signs on as a Platinum member. Its engineers sit among the protocol’s core maintainers and contributed Tasks and Elicitations.
May 2026: the managed AWS MCP Server goes GA. One endpoint, “15,000+ AWS API operations” across 300+ services, IAM and SigV4 as the boundary, CloudTrail as the audit log, no charge. The tool list? Nine tools: documentation search, skills, regions, presigned URLs, task polling, and two ways to act.
July 15, 2026: AWS deprecates one of those two, call_aws, with removal scheduled for August 31. The recommendation is run_script, which “execute[s] Python code in a sandboxed environment with AWS API access,” server-side, with inherited IAM permissions and no network, for “parallel API calls, multi-step workflows, cross-service checks, and retry logic.”
Read that sequence as an architect. The company that invented API-first now exposes its entire API to agents through a single tool that runs a program.
Why the shape was forced
Scale made the decision. At the roughly 950 tokens per definition implied by Anthropic’s own measurements, fifteen thousand tools would be fourteen million tokens of catalogue before the first turn. The flat menu was never an option for Amazon. It is only an option for the rest of us because our APIs are small enough to hide the bill.
So AWS arrived, by necessity, at the three patches I catalogued earlier this month in We Fixed the Engine and Threw Away the Map: a uniform interface, code execution instead of tool picking (Cloudflare’s Code Mode, Anthropic’s code execution with MCP), and tool search. The last one is a product: AgentCore Gateway puts your Lambdas, APIs and MCP servers behind one endpoint and ships a semantic search tool that is “particularly useful when you have many tools.”
Call this what it is. Not MCP-first. API-first, with MCP as the socket and a sandbox as the client. I have written that MCP won where the Semantic Web lost because a junior developer can write a server in an afternoon. Amazon’s year shows the second half of that sentence: at scale, the friction comes back as tokens, and the cure is the fundamentals the afternoon skipped.
Half the map
Now the part Amazon has not done.
Discovery in the AWS server is still search. The agent finds the right operation through search_documentation, through “skills” (workflows written in prose, formerly Agent SOPs), and through whatever it remembers of the CLI. Nothing in any response says: from here, these are the legal moves. Order and preconditions, role before function, VPC before subnet, are the model’s problem, exactly as they were with a flat catalogue, only now inside a script.
IAM tells you a call was illegal after the round trip. A pruned choice costs nothing; an AccessDenied costs a full turn. And state lives in the model’s context, or in the script’s variables, never in the representation. That is the inversion of hypermedia I described in September, and the 2026-07-28 specification still forbids the tool list from varying with state.
The one hopeful sign is in the MCP roadmap of August 22: the maintainers want a server to “offer a small entry point and reveal more.” Those are the bones of hypermedia. Whether it becomes a map depends on one question. Does “more” depend on the state you are in, or only on the query you typed? The first is Fielding. The second is pagination.
What MCP-first should mean
If you are deciding your integration strategy this quarter, copy Amazon’s shape and then finish it.
Keep MCP for what it is good at: transport, auth, client ubiquity. Do not wrap endpoints in tools; expose one entry point and let agents write code against a typed API. Put policy at the boundary, as AWS did with IAM. Then add the half AWS has not, the state machine and shared vocabulary I proposed in April: representations that carry the current state, the affordances valid in it, and the preconditions still unmet. Three tools, enter, follow, describe, and the server prunes the tree. The design is in the September essay; the schema fits on a page.
Measure one number: tokens per completed task. Not per call. Per task, retries included. The forty-five CIOs and CTOs who told me agentic drift worries them more than hallucination were describing the absence of choreography, and no gateway search will fix that. As I argued in The 100x Agent Illusion, you engineer the system, not just the code.
The memo, revised
The Bezos memo never said “expose everything through a chatbot.” It said service interfaces, designed to be externalizable. The 2026 version is one clause longer: interfaces that tell the agent what is legal now.
Ten years ago my thesis argued that the server should “guide the client into the map of its entry points.” Amazon has built the best engine room in the industry: one socket, a sandbox, a policy wall. The map is still ours to draw.
Sources and further reading
Primary sources: AWS API MCP Server, July 2025 · The AWS MCP Server is now generally available, May 2026 · Understanding the MCP Server tools (Agent Toolkit for AWS docs) · AWS MCP Server reaches GA (InfoQ) · How AgentCore Gateway supports the MCP 2026-07-28 spec · AgentCore Gateway semantic tool search · As agentic AI explodes, Amazon doubles down on MCP (The New Stack) · Linux Foundation announces the Agentic AI Foundation · MCP specification 2026-07-28: Tools and changelog · The New MCP Roadmap, August 2026 · Anthropic: Advanced tool use and Code execution with MCP · Cloudflare: Code Mode · Steve Yegge’s platform rant (the Bezos mandate) · Fielding: REST APIs must be hypertext-driven
Earlier on APILama: We Fixed the Engine and Threw Away the Map (Sept 2026) · Upgrading the Model Context Protocol (Apr 2026) · The AI ROI Reality Check (Apr 2026) · The 100x Agent Illusion (Apr 2026) · From DeepGraphs to MCP (Jan 2026) · Uniform Interfaces in a Nutshell (2018) · Adding Rules on Existing Hypermedia APIs, WWW 2015 · hub: The Agentic Web & MCP
Leave a Reply