Standards · ~10 min read
The single most important thing that happened in AI infrastructure in the last 18 months wasn't a model release. It was a protocol. The Model Context Protocol — MCP — is the first real attempt at a universal connector between AI clients and the data, tools, and memory they need. And it quietly changes the answer to the question most people don't ask out loud: who owns the version of you the AI knows?
One-sentence answer: MCP is an open JSON-RPC protocol, released by Anthropic in November 2024, that lets any AI client talk to any data source or tool through a shared interface — N clients × M tools collapses to N+M.
The shape is borrowed almost directly from the Language Server Protocol (LSP). LSP solved the same combinatorial problem for code editors in 2016: instead of every editor writing its own integration with every language, editors speak LSP and language servers speak LSP, and the matrix flattens. MCP does that for AI clients and context sources — file systems, databases, calendars, vector stores, memory vaults, web tools.
The protocol defines three primitives a server can expose: tools (things the model can call), resources (things the model can read), and prompts (reusable prompt templates the client can offer). Transport is pluggable: stdio for local servers, HTTP/SSE for remote ones.
MCP isn't a thought experiment. The adoption curve is the steepest any AI infrastructure standard has had:
For context: it took LSP ~3 years to reach broad multi-editor support. MCP did it in ~12 months. The reason isn't mysterious — the AI client market is small (a handful of major clients), the SDKs are good, and every client team wanted the integration matrix to stop being their problem.
The clearest way to see what MCP changes is to draw the picture three times.
Before MCP. Every AI client you used had its own integration story. ChatGPT plugins lived in OpenAI's plugin store; Claude had no plugins at all; Cursor had its own extensions; agent frameworks each defined their own tool schemas. Switching clients meant rebuilding integrations.
MCP. One protocol, one handshake, one schema for tools and resources. Write a server once, every MCP-aware client speaks to it.
The deeper unlock. Once the protocol exists, the same logic applies to memory: a single memory source you control, read by whatever client you're using that day. The provider stops being the owner of "the version of you the AI knows." That's the part that matters for the rest of this site.
One-sentence answer: MCP is plumbing — it moves bytes between client and server, it doesn't tell you what those bytes should be.
A few things people expect from MCP that it doesn't deliver:
Pick the analogy you like. MCP is to AI clients what USB-C is to devices, what LSP was to code editors, what SMTP was to email. The common thread is that none of them are exciting on the day they ship. They become important when the things on either side of them stop having to know about each other.
The practical effect, two years from now, is that the question "which AI do you use?" becomes less interesting than "where does your memory live?" The first question gets a different answer every six months. The second question, if you set it up well, gets the same answer for a decade.
Konshus is an MCP server you control, with the memory layer already built. Import your ChatGPT export, your Claude history, your journals and docs; the vault distills them into atoms (decisions, preferences, patterns, voice). Connect Claude Desktop or any MCP-aware client to /api/mcp and that client now reads from your vault — without us touching its chat traffic or its account. Revoke the token at any time. See the connection flow →
If you decide a folder of Markdown files and a hand-rolled MCP server does the job, that's a reasonable answer too. The point of MCP is that you get the choice.
Konshus is the memory layer underneath. Import once, connect via MCP from Claude Desktop, ChatGPT, or whatever ships next. Encrypted, exportable, never used for training.
Meet Konshus