Skip to main content

Connectors / reconnect (MCP Servers via http)

When using connectors, I get issues on a daily basis with connecting to MCP servers via connectors for example, Bright Data.

Some of them have connection timeouts and disconnect you. The problem is that when the LLM tries to use them, it can't, because the connector is disconnected. The only way to fix it is to go to Settings / Connectors / Reconnect.

It would be really great to have a connection health check with a reconnect on each LLM call so it can automatically reconnect.

Status: Completed4 comments

Log in to comment and vote

Comments4

  • Daniel Nguyen changed status to In QA
    Team•

    Jul 19

    Pinned

    I’ve fixed this in v2.14, can you confirm?

    • David

      •

      Jul 19

      so far so good… but will take more time to fully confirm it…

      will report if there are still any issues

  • Daniel Nguyen

    Team•

    Jun 25

    This turns out to be a pretty challenging problem. It used to work like that, relying on the health check (ping) from MCP but then many servers never honor it. Some don’t even implement it and would down right rate limit our client.

    So in the recent versions, I stopped using ping/health check altogether and only connect on demand.

    The problem is that sometimes it wasn’t able to reconnect fast enough for the LLM, or ran into an issue and after a few retries, just stopped to avoid being banned / draining resources.

    I think one solution now is probably to expose a custom tool to reconnect/connect a certain MCP server — kinda meta but LLM is now smart enough to call it. It might be better than any health check impl tbh.

    Quick question though, when you go to Settings > Connectors > Reconnect, does it prompt for browser auth again? If yes, then it’s a bug where BoltAI might not persist the client creds correctly.

    • David

      •

      Jun 25

      Ah… I see.

      I don't use OAuth but a bearer token, so I never get prompted again. I think a custom tool to let an LLM reconnect/connect a bridge would probably be really good.

      I noticed my LLM does multiple retries on bad MCP servers anyway. Not sure if you could auto-reconnect based on a bad response? You would probably be too late, I guess, since the LLM does that within the same chat session. So your idea to give the LLM the option to reconnect is probably best.

      But you must make sure that it can only reconnect the allowed MCP servers, so that it can't go rogue and reconnect any MCP server I have configured. That being said, sometimes it also happens that the MCP server disappears from the list of MCP servers in the chat dropdown (you could maybe cache the list (so LLM knows it has that tool) and allow it to reconnect that as well).