A couple of questions to help debug. What transport type is your Bugsnag MCP using, HTTP or stdio? Stdio servers run inside the VM and depend on the environment. HTTP servers are proxied through the backend. This can affect how Automations see MCP.
Also, did you enable the MCP tool when creating the automation, in the list of available tools?
Let me know the details above so we can narrow it down.
Ah, I didn’t realize that you needed to pick-and-choose the MCPs that the Automation would have access to. Was expecting that would have access to all of them by default.
Will try that first, and can come back with more details if it doesn’t work.
@mohitjain I would also note that it seems like the same issue happens when an agent is invoked via tagging it on Slack.
I am not sure if the root cause issue is the same, or it’s a separate issue.
A few different issues are getting mixed together in this thread, so let me address them individually.
@evsasse – Glad to hear the initial issue was likely resolved by adding the MCP directly to the automation’s tool list. For anyone landing here: MCPs configured at the account or team level don’t automatically carry over to automations. You need to explicitly add them via the “Add Tool or MCP” option when editing your automation.
@erezro – Thanks for sharing the workaround of passing credentials directly to the agent via secrets. The underlying issue you hit (MCP tools added to the automation but the agent still can’t find them at runtime) is a known intermittent problem that the team is actively working on.
@petar-studocu – Your configuration looks correct. The Sentry and Shortcut MCPs are properly added to the automation’s tool list with the right stdio commands and environment variables. The fact that it works when you trigger an agent via the “new agent” web UI but fails when triggered via Slack or as an automation is a known issue with how MCP resolution works differently across these trigger paths. The team is aware and tracking this.
For anyone experiencing MCP tools not being found despite being configured correctly, here are a few things worth knowing:
This is a known intermittent issue. MCP tool discovery in automations can fail transiently, particularly when triggered via Slack or other automation triggers (as opposed to manually launching an agent via the web UI). The team is working on improving reliability here.
Configure MCPs via the dashboard, not .cursor/mcp.json. Cloud Agents and Automations do not read from .cursor/mcp.json in your repo. Use:
https://cursor.com/agents (when editing your automation)
Hardcode secret values instead of ${env:...} interpolation. If you’re using environment variable interpolation syntax in the MCP config, note that ${env:VAR_NAME} does not currently resolve in Cloud Agent/Automation MCP configs. Enter the actual values directly – they are encrypted at rest.
Workaround: pass credentials as agent secrets. As @erezro noted, if MCP access remains unreliable, you can bypass it by sharing credentials directly with the agent via the Secrets/Environment configuration and having the agent make API calls directly.
We’ll update this thread when there’s progress on the reliability improvements.
Thanks for the clarification, @mohitjain
Eagerly awaiting an update from the team, then.
I would just add that even when accessing a cloud environment agent from the cursor dashboard (home page), and selecting the MCPs as tools there, having added them on the https://cursor.com/dashboard/integrations page, it still may NOT work.
I attempted debugging this, but some colleagues report it works, and indeed for me it seems intermitted.
Do let me know if this is the same issue or yet a separate one.
Thanks for the screenshots – this is really helpful.
You’re right that this isn’t limited to automations or Slack-triggered runs. MCP tools can also fail to load in manual cloud agent sessions started from the dashboard. Our team has identified what’s causing this and is actively working on a fix.
In the meantime, if MCP tools don’t appear in a session, starting a new one sometimes resolves it. Configuring your MCP servers and secrets directly in the Cursor Settings dashboard is also more reliable.
Is there any update on this? I desperately want to build a ton of stuff around cloud agents / automations and throw a lot of money at you guys, but the MCP functionality seems so broken that it’s making it impossible. There’s this too: Cloud Agents + MCP - #20 by Colin
Same problem here. I want to use grafana mcp on cloud agents to automate incident investigation and response and it is not working. The other http mcp servers are working, but not this one with stdio.
