MCP is one of the greatest API adoptions ever
Nov 18, 2026 · 4 min read · #ai · #mcp · #apis · #graphql
Anthropic put out MCP (Model Context Protocol) in November 2024. Two years later pretty much every SaaS I use has an MCP server, and so do all my own products. I’ve been around for SOAP, REST and GraphQL, and I can’t remember a standard going this fast.
Some numbers:
- The TypeScript and Python SDKs went from ~2 million to ~97 million downloads a month in about 16 months.
- In December 2025 it moved to the Linux Foundation’s Agentic AI Foundation, with OpenAI, Google and Microsoft on board. GraphQL took 3.5 years to get there.
- In Postman’s 2025 survey 70% of developers knew MCP, less than a year after launch. GraphQL is at 33% usage after ten years.
To be fair, only 10% used MCP regularly at that point. Everyone knows it, not everyone uses it yet.
Why it worked
- It’s dumb simple. It’s JSON-RPC. The server lists its tools with a name, a description and a JSON schema, the client calls them. You can build one in an afternoon.
- The docs are the API. The model reads the description and figures out when to call what. No integration code.
- SaaS companies get something out of it. One server and you’re inside Claude, ChatGPT, Copilot, Cursor and whatever comes next month. Compare that to GraphQL, where the backend team does the work and the frontend team gets the benefit.
- Everyone wants to talk to their SaaS from chat. “Send this survey to the project team.” Done.
It’s just an API
Most people think MCP is an agent thing. It isn’t really. It’s a list of things your service can do, with typed inputs and OAuth, that any client can discover. That’s an API.
So I use it like one, for plain server to server calls. My apps call each other’s MCP servers to push and pull data, no LLM in between. Scripts and cron jobs too. Same auth, same tool list, same client library. No separate REST API to build and maintain next to it. Agents just come along for free.
What I’d like: binary
The one thing I miss is a simple way to push and pull files. Right now it’s all JSON, so binary goes in as base64. That’s a third bigger, falls over around 4MB, and when a model has to copy the base64 into a tool call it happily corrupts it.
The workaround everyone uses: a tool hands back a presigned URL, and you do a normal HTTP upload or download next to MCP. Works, but now every server invents its own flavour of it.
And often it doesn’t even work. Agents mostly run in a sandbox, and the sandbox only lets the MCP connection out, not some random bucket URL. So the agent gets a presigned URL it can’t reach. If uploads are part of the spec, the client knows what that URL is for, and sandboxes can allow exactly those uploads instead of opening up the whole network.
What I want is boring HTTP. Let a tool say “this input is a file”, give the client an upload URL, PUT the bytes, done. Same for downloads.
And good news, that’s roughly where it’s going. The file uploads working group just merged a first draft of an io.modelcontextprotocol/files extension:
- A tool marks a file argument with
x-mcp-filein its input schema. - Small files can still go inline as base64, up to the server’s
maxInlineBytes. - For bigger ones the client first calls the tool with just the metadata. The server checks quota and permissions and answers
upload_requiredwith upload targets. - The client sends the bytes there with a plain
PUT,POSTor multipart form, and retries the call. - Files coming back are download targets in the result.
The targets are meant to match what S3, Azure Blob, Cloudflare and friends already hand out, so it’s basically the presigned URL trick, standardized. The model sees one tool call, the bytes and credentials never go through its context.
It’s experimental and the comments on the PR are still going at it (HTTP PUT vs a new MCP method, expiry, retries), but the direction is right. Hope they keep it as dumb as the rest.
It’s not all roses
- Security is a mess. No identity, no least privilege, no audit trail in the protocol itself, and 30+ CVEs in the first two months of 2026.
- Tool descriptions eat your context. Three servers and you’ve burned 20k+ tokens before you’ve asked anything.
- Lots of servers are lazy REST wrappers with a hundred tools.
Still, those are the problems of something that won. REST had them too.
Fun fact: back in 2019 I wrote about GraphQL protocols, standard agreements on top of an API so clients don’t need custom work for every service. MCP is basically that, minus the GraphQL. Guess the trick was to keep it boring.
enjoyed this? there's 21 years more where that came from
More writing →