The Background Agent always has MCP disabled, even though they are configured in the project. My use-case:
Sentry errors come to me via slack, and the Sentry AI writes very nice summaries of the issue by watching replays and looking through all the logs. I can only access the output of the Sentry AI via the Sentry MCP. The analysis from the Sentry AI is enough information for a Background Agent to do a very good job at resolving the issue, but the background agent doesn’t start with the MCP enabled. I can open the agent, go to settings, and turn it on and it will work.
If it started with the MCP enabled, then I could trigger the background agent from Slack, inline with the sentry issue notification. It could all run nicely in the background and be ready for me to review when done.