What actually deserves an MCP slot in your Claude Code or Cursor setup

Every MCP server you add to Claude Code or Cursor costs something — a slightly slower startup, a slightly longer tool list for the model to reason over, one more thing to keep an API key alive for. Past four or five servers, most engineers start being deliberate about what earns a spot. That's the right instinct, and it's worth being explicit about the bar rather than just accumulating whatever looked interesting on a given week.

A reasonable bar for a memory server specifically

"Memory" MCP servers are common enough now that it's worth asking what kind of memory you're actually adding, because the category splits in two:

• Per-developer memory — your own preferences, your own past conversations, personalized to you specifically. Useful, but it stays on your machine, in your head, in your config. A teammate gets none of it.
• Team memory — the decision your teammate's agent made last Tuesday, the failed migration approach someone already tried, the "we tried that, it broke prod" that would otherwise live only in one person's terminal history. This kind only has value if it's shared.

If you only have room for one memory server, the question is which category actually saves your team time: memory that improves how your agent talks to you, or memory that stops your team's agents from re-solving the same repo problem three times across three people.

The zero-friction config, either way

Whatever you add, it should cost you nothing beyond a copy-paste. threadctx's local mode needs no signup or key at all — it's a free, on-disk store you can try in thirty seconds:

The moment a second teammate's agent needs to see the same memory — in Claude Code, Cursor, or anything else that speaks MCP — that's the point where it's worth switching on Team mode instead of writing it off as "another memory plugin."