Request: Make it possible to enable auto-run mode just for a specific set of MCPs.
Loving Cursor so far, particularly with what the Agent mode is capable of with my MCP tools to access internal databases + Cursor rules for instructions. However, it adds incredible friction to the Agent workflow to need to supervise it when it needs to make a handful of MCP calls. My organization has disabled auto-run mode for safety reasons, however, the specific MCP server I run is read-only. Therefore I think it’s needless that we can only disable auto-run mode for all cases. I would love it if our organization could specifically whitelist MCP servers it knows to be safe for auto-run mode. It sounds trivial but being able to kick off agent mode without needing to approve every MCP call would save me and my team members a tremendous amount of time.
Right now, enabling auto-run mode for MCP is all-or-nothing and enables it for all MCPs. This means that our org has disabled certain MCP URLs, because we require auto-run mode for some MCPs.
We need to consider universally disabling auto-run mode for MCPs so that we can enable more in-development servers, which is a huge bummer.
Would ask the team to please consider building out more per-server controls.
@loubacker I want to call out here that the linked thread describes a related but separate issue from this. The linked thread was resolved, because Cursor separated auto-run for MCP from auto-run for commands. It doesn’t address my concern, which is separating auto-run at the level of granularity of multiple MCP servers.
I want to emphasize the real need for a fully independent “Auto-Run MCP” control, separate from the global “Auto-Run Mode”. Right now, when I disable “Auto-Run Mode”, all automations are turned off (MCPs included). And when I enable the global auto-run just to allow MCPs, I end up bringing along every other automated behaviors (command execution, file writes, linters, etc..). So, the only related option today is the “MCP Tools Protection” checkbox, but it only blocks MCPs once the general auto-run is active. In other words, it doesn’t solve our use case of wanting to keep only MCP auto-run enabled without activating full auto-run.
What you raised is exactly the point. I think we need a standalone “Auto-Run MCP” toggle outside of “Enable auto-run mode,” with the ability to whitelist by URL or server.
While I agree with you, what I’m proposing is that we have different settings for “Auto-run MCP Server A”, “Auto-run MCP Server B”, and “Auto-run commands” assuming our project utilizes two servers, A and B.
To clarify for the thread, the Auto-run feature I requested has not already been implemented. My organization sets the global auto-run behavior for the entire org, and we don’t have MCP-level granularity options for this.
Further, I don’t think the shared screenshot reflects Auto-run behavior, only selecting enabled MCPs for an individual agent run.
I suggest implementing a per-tool allow list. Currently, Cursor has a allow list command for auto-run. Enabling a allow list for each tool in MCP would be great.
