AI
Why LLMs Need MCP to Act on Real Infrastructure
LLMs take text in and return text out. They cannot call Argo CD, read logs, or trigger a sync on their own. The fix is a thin layer: an AI client sends a structured JSON-RPC request to an MCP server, which calls the third-party service and translates the response back into text the model can use. MCP defines two transport modes. Run it locally and use stdio. Deploy it next to your service and use HTTP, either server-sent events or HTTP/2. Authentication passes through to the upstream service via token.
What the MCP for Argo CD Server Exposes
The project lives under Argo Project Labs on GitHub and its CNCF Slack channel had 16 members at the time of the talk. The server implements tools only, no resources. Coverage is one-to-one with the Argo CD CLI and UI: sync applications, inspect Kubernetes manifests, stream live logs, and query resource state. Prompts describe each tool so the LLM knows when to call it. The happy path is complete and running in production at Intuit. Edge cases such as applications in non-default namespaces are still in progress.
AI Agents Handling Sync, Bulk Creation, and Automatic Rollback
Three use cases were shown. First, creating a single application from a free-text prompt specifying a Git repo, branch, and target cluster. Second, scanning all directories in a repository and creating one Argo CD application per directory. A one-sentence prompt produced results that would take a human noticeably longer to complete manually. Third, a pre-recorded demo showed an agent deploying a new image version, monitoring application health, detecting degradation, and rolling back the Git manifest to the previous version without human intervention.
Intuit’s Scale and Why the Argo CD Extension Failed First
Intuit runs 350 Kubernetes clusters, 3,000 production services, and 50,000 namespaces. The team built an Argo CD UI extension that pulled logs, Kubernetes events, live state, and desired state, fed them to an LLM, and returned a diagnosis. It correctly identified an invalid image tag and a pod template that exceeded namespace memory limits. Adoption failed anyway. Engineers already familiar with Kubernetes went straight to logs. The users who needed help most were the ones posting in Slack support channels, not browsing the Argo CD UI.
The Support-Channel Bot: MCP Closes Tickets in Minutes
Intuit runs more than 40 Argo CD instances. They deployed the MCP server as a reverse proxy, giving a single service access to all 40 instances and one security entry point. The bot in their Slack support channel connects to that MCP service. In one logged conversation, a developer posted a production deploy failure with an Argo CD link. One minute later the bot identified two separate failures. Two and a half minutes after that it traced the root cause: the application was pointed at an E2E environment URL instead of the production Spring Cloud Config server URL.
Agent Skills: Codifying the Troubleshooting Recipe Once
The same troubleshooting logic now lives in at least two places: the original UI extension and the Slack bot. Adding a third use case means copying it again. Intuit is experimenting with agent skills, markdown files that instruct an agent on exactly which steps to run when an application is degraded. The skill extracts the Argo CD base URL from context, passes it as a header to the MCP service, and follows a defined sequence to pull logs and events. Combining agent skills with the MCP service is the direction they are moving, though the team flagged it as a topic for a future talk.
Notable Quotes
it’s kind of boring and not so useful. Alexander Matyushentsev · ▶ 01:56
I was too scared to rely on um AI to execute it perfectly so I just pre-recorded it. Alexander Matyushentsev · ▶ 12:18
to me is you know convincing enough that AI is not just a toy. It can be used to save time of engineers. Alexander Matyushentsev · ▶ 12:55
I was personally very very well amazed. I thought it was it would be a big hit and it would be a successful project. Unfortunately it wasn’t. Leonardo Luz Almeida · ▶ 19:14
I don’t see one but two different failures in your application. Leonardo Luz Almeida · ▶ 24:55
Key Takeaways
- MCP for Argo CD maps every CLI operation to a tool, making the full API available to any AI agent.
- Intuit’s 40-instance deployment uses MCP as a reverse proxy with a single security entry point.
- Adoption of AI tooling depends on meeting engineers where they already work, inside Slack.
About the Speakers
Alexander Matyushentsev is co-creator and lead maintainer of Argo CD at Akuity, where he serves as Chief Software Architect. He has spent nearly a decade building Argo CD, Argo Rollouts, and Argo Workflows, and is a core contributor to the broader Argoproj ecosystem.
Leonardo Luz Almeida is a Staff Software Developer at Intuit and an active Argo maintainer. He leads Argo CD and Argo Rollouts operations across Intuit’s production fleet and splits his time between internal platform work and open source contributions.