Why has my MCP connection stopped responding?
1. The rate limit is reached
This is the most frequent cause. Every token has two ceilings:
- 30 calls per minute, to absorb a burst;
- 400 calls per hour, for sustained use.
An agent looping over your memory hits the first in seconds. There is nothing to repair: wait, and the counter drains by itself. A minute is usually enough.
The ceilings are per token, not per person: two AI clients wired with the same token share the same allowance. If you connect both ChatGPT and Claude, give each its own token.
2. The token was revoked or replaced
A revoked token stops answering immediately, and a regenerated token does not replace the old one inside your client: you have to paste the new one in.
- Open Integrations, MCP tab.
- Check that the token you are using is still listed.
- If you generate a new one, copy it straight away: it is shown only once.
- Replace it in your AI client's configuration, then restart the client.
3. The client is misconfigured
Every AI client has its own procedure, and they do not resemble each other. The Integrations page asks you to pick the client first (Claude, ChatGPT, Cursor, Windsurf, other), then shows the matching configuration, ready to copy.
ChatGPT is the longest of the five: connecting goes through its developer mode, which adds several steps compared to the others. That is not a fault in your setup.
Didn't work?
- It worked yesterday and not today, with nothing changed. Look at the rate limit first: an agent left looping eats the hourly ceiling without anyone noticing.
- The client sees no tools at all. The configuration is not being picked up: in almost every case the client was not restarted after the change.
- Tools appear but every answer is empty. Check that the token points at the right memory, and that the memory has something in it.
FAQ
How do I tell a rate limit from a token problem?
A rate limit clears itself: try again a minute later. If access returns without you touching anything, it was the rate limit.
Can the ceiling be raised?
These ceilings protect your memory as much as the service: without them, a looping agent can consume without end. If legitimate use exceeds them, write to us rather than multiplying tokens.
Does one token reach all my memories?
No. A token is tied to one memory and one scope, read or read and write.