What is an MCP server?

The short answer

MCP (Model Context Protocol) is an open standard, introduced by Anthropic in late 2024, that lets AI agents call external tools — read data, write data, hit an API — through one common interface, instead of a bespoke integration for every agent-tool pair. An MCP server is a small program that exposes a set of tools (each with a name, a description, and a typed input schema) over that protocol. Any MCP-compatible client — Claude Code, Cursor, and a growing list of others — can connect to it and call those tools.

Why it exists

Before MCP, giving an AI agent a new capability — search the web, query a database, remember something — meant writing a custom plugin for each specific agent product. MCP standardizes the wire format: build one server, and every MCP-compatible client can use it immediately. It's the same shape of idea as USB-C for hardware or LSP for editors — one interface, many implementations on both sides.

How a client uses an MCP server

1. The client launches or connects to the server (commonly over stdio, or HTTP for remote servers).
2. It asks the server what tools are available — each tool has a description written for the model, so the agent knows when to call it.
3. During a task, the model decides to call a tool (e.g. memory_query), the client executes it against the server, and the result is fed back into the model's context.

This is why tool descriptions matter so much — MCP is pull-based, so the model has to be prompted by the description to actually call a tool. A vague description means an agent that never uses the capability you built.

How threadctx uses MCP

threadctx is an MCP server with exactly two tools: memory_write(content, tags?), which an agent calls after resolving something non-obvious, and memory_query(task_description, max_results?), which an agent calls before starting risky or repeated work. In local mode the server stores memory in a plain JSON file on disk; in cloud (Team) mode it stores memory centrally, shared across everyone on the repo, with real semantic search. Since pull-based tools only get called if something prompts them to, threadctx sends its "check memory before, write after" instruction directly in the MCP handshake and auto-adds it to your project rules on first start — not just a tool description and a hope. Because it's MCP, the identical server and config work in Claude Code and Cursor with zero client-specific work.

Try it

Free, local, no signup — this runs the threadctx MCP server on your machine right now.