Today’s AI coding agents are impressive. They can generate complex multi-line blocks of code, refactor according to internal style, explain their reasoning in plain English, and more. However, AI agents will take you only so far unless they also can interface with modern devops tools.
This is where the Model Context Protocol (MCP) comes in. MCP is a proposed universal standar…
Today’s AI coding agents are impressive. They can generate complex multi-line blocks of code, refactor according to internal style, explain their reasoning in plain English, and more. However, AI agents will take you only so far unless they also can interface with modern devops tools.
This is where the Model Context Protocol (MCP) comes in. MCP is a proposed universal standard for connecting AI assistants with external tools and data. Interest has heated up since the protocol’s debut in late November 2024, with major tech companies rallying MCP support within new releases, alongside strong community interest.
For devops, MCP gives AI agents new abilities across common operations: Git version control, continuous integration and delivery (CI/CD), infrastructure as code (IaC), observability, accessing documentation, and more. By linking natural language commands to multi-step, back-end processes, MCP essentially enables “chatops 2.0.”
Below, we’ll explore official MCP servers that have emerged across popular devops tools and platforms, offering a cross-section of servers that cater to different devops capabilities. Most are straightforward to configure and authorize within MCP-compatible, AI-assisted development tools that support remote servers, like Claude Code, GitHub Copilot, Cursor, or Windsurf.
GitHub MCP server
It’s rare to meet a developer who doesn’t use GitHub in some form or fashion. As such, GitHub’s official MCP server is quickly becoming a popular way for AI agents to interact with code repositories.
GitHub’s remote MCP server exposes a range of tools that let agents perform repository operations, create or comment on issues, open or merge pull requests, and retrieve project metadata on collaborators, commits, or security advisories.
It also includes endpoints for CI/CD management through GitHub Actions. For example, a command like “cancel the current running action” could invoke the cancel_workflow_run tool within the GitHub Actions tool set.
Compared to other MCP servers, GitHub’s server offers unusually rich capabilities that mirror the APIs of the GitHub platform. However, for safety, you can always configure a --read-only flag to prevent agents from performing mutations.
Notion MCP server
Although not strictly devops at its core, Notion has become commonplace for team visibility across disciplines. For devops, the official Notion MCP server can help agents surface relevant notes and process documentation.
For instance, you could instruct an agent to reference internal style guides or operational runbooks stored in Notion, or issue a command like “Add a page titled ‘MCP servers we use’ under the page ‘DevOps’,” which would trigger a corresponding action through Notion’s API.
You can call Notion’s remote MCP server from your IDE, or build it locally and run it using the official Docker image. Notion’s MCP can be treated as a low-risk server as it has configurable scopes and tokens for managing Notion pages and blocks.
Atlassian Remote MCP server
Another interesting MCP server is the Atlassian Remote MCP server, which connects IDEs or AI agent platforms with Atlassian Cloud products such as Jira, the project management tool, and Confluence, the collaboration platform.
Atlassian’s MCP server, documented here, lets external AI tools interface with Jira to create, summarize, or update issues. It can also retrieve or reference Confluence pages and chain together related actions through the MCP client, like retrieving documentation from Confluence before updating a linked Jira issue.
You could imagine telling an agent, “Update my Jira issue on user testing for the payments app based on this latest bug report,” and pointing it to relevant logs. The server would then handle the update within Jira.
Currently in beta and available only to Atlassian Cloud customers, the Atlassian MCP server supports many MCP-compatible clients and uses OAuth 2.1 authorization for secure access.
Argo CD MCP server
The Argo CD MCP server is developed by Akuity, the original creators of Argo CD, the popular open-source CI/CD tool that powers many Kubernetes-native GitOps workflows. The MCP server wraps calls to the Argo CD API, and provides tools that allow users of AI assistants to interact with Argo CD in natural language.
Akuity’s MCP server has two main tools for applications (the deployments Argo CD manages) and resources (the underlying Kubernetes objects). The application management tool lets agents retrieve application information, create and delete applications, and perform other operations. The resource management tool allows agents to retrieve resource information, logs, and events for specific applications, and run actions on specific resources.
Using the Argo CD MCP server, you can do a lot of the same things you’d typically do in the Argo CD UI or CLI, but driven by natural language. For example, Akuity shares sample prompts such as “Show me the resource tree for guestbook” or “Sync the staging app.”
For such commands to work, you’ll need to integrate the Argo CD MCP server and have access to a running Argo CD instance with the proper credentials configured.
Lastly, although Argo CD is a popular choice, it’s not the only widely used CI/CD tool. Jenkins users may be interested to know that there is a community-maintained MCP Server Plugin for Jenkins.
Grafana MCP server
Grafana, the popular data visualization and monitoring tool, is a mainstay among devops and site reliability teams. Using the official MCP server for Grafana, agents can surface observability data to inform development and operations workflows.
The Grafana MCP server lets agents query full or partial details from dashboards, which combine system performance metrics and health data monitoring from various sources. It can also fetch information on data sources, query other monitoring systems, incident details, and more.
The tool set is configurable, so you can choose what permissions the agent has. Plus, Grafana has optimized how the MCP server structures responses to minimize context window usage and reduce runaway token costs.
For example, an MCP client might call the get_dashboard_property tool to retrieve a specific portion of a dashboard by its UID.
Terraform MCP server
Although alternatives have emerged, HashiCorp’s Terraform remains a leading choice for infrastructure as code. That makes its official MCP server an intriguing option for AI agents to generate and manage Terraform configurations.
The Terraform MCP server integrates with both the Terraform Registry APIs and Terraform Enterprise/HCP services, allowing agents to query module and provider metadata, inspect workspace states, and trigger runs with human approval. It also exposes Terraform resources such as runs, registries, providers, policies, modules, variables, and workspaces.
For example, a command like “generate Terraform code for a new run” could use the create_run operation, after which the agent might validate and plan the configuration before applying it.
The Terraform MCP server ships with an AGENTS.md file, which acts as a readme for agents to interpret tools. At the time of writing, the Terraform MCP is intended only for local use, rather than remote or hosted deployments.
Alternatively, if you’re using OpenTofu for IaC, consider checking out the OpenTofu MCP server. Some advantages of OpenTofu’s MCP are that it can be run locally or deployed in the cloud, it’s globally distributed on Cloudflare Workers, and it’s 100% open source.
GitLab MCP server
Another Git version control and devops platform is GitLab, which offers an MCP server for its Premium and Ultimate customers. The GitLab MCP server, currently in beta, enables AI agents to gather project information and perform operations on GitLab APIs in a secure way.
The GitLab MCP server allows some state changes, such as creating issues or merge requests. The other functions are mainly for data retrieval: retrieving information on issues, merge requests, commits, diffs, and pipeline information. It also includes a general search tool, which can handle a request like “Search issues for ‘failed test’ across GitLab.”
GitLab’s MCP documentation is thorough, with plenty of sample natural language expressions that the MCP server can satisfy. The server supports OAuth 2.0 Dynamic Client Registration.
Snyk MCP server
Snyk, maker of the Snyk security platform for developers, provides an MCP server with the ability to scan and fix vulnerabilities in code, open source dependencies, IaC code, containers, and software bill of materials (SBOM) files. It also supports creating an AI bill of materials (AIBOM) and other security-related operations.
For AI-assisted devsecops, integrating the Snyk MCP server could let an agent automatically run security scans as part of a CI/CD workflow. These scans can even be orchestrated across other MCP servers, like fetching repository details via the GitHub MCP server before initiating a Snyk scan.
A prompt like “Scan the repo ‘Authentication Microservice’ for security vulns” could instruct an agent to locate the repository using GitHub MCP, then invoke Snyk tools such as snyk_sca_scan or snyk_code_scan to identify known vulnerabilities, injection flaws, leaked credentials, and other risks.
The Snyk MCP server runs locally and uses the Snyk CLI to execute these commands through authenticated API calls. Snyk does not offer a hosted, remote version of the MCP server.
AWS MCP servers
The cloud hyperscalers have worked quickly to release MCP servers that integrate with their ecosystems. AWS, for instance, has rolled out dozens of specialized AWS MCP servers to allow AI agents to interact with all manner of AWS services. Some are provided as fully managed services by AWS, while others can be run locally.
For instance, the Lambda Tool MCP server allows agents to list and invoke Lambda functions, while the AWS S3 Tables MCP server could be used by an agent to query S3 table buckets or create new S3 tables from CSV files. The AWS Knowledge MCP server connects agents with all of the latest AWS documentation, API references, and architectural guidance.
A query to this knowledge server, like “pull up the API reference for AWS’s managed Prometheus tool” would correspond with the correct up-to-date information, optimized for agentic consumption.
Users of Microsoft Azure might want to evaluate the Azure DevOps MCP server. Other clouds, like Alibaba, Cloudflare, and Google, are currently experimenting with MCP servers as well.
Pulumi MCP server
Pulumi, another popular option for IaC, has also launched an official MCP server. The MCP server allows agents to query a Pulumi organization’s registry, which provides access to cloud resources and infrastructure, and execute Pulumi commands.
For example, in this walk-through, Pulumi shows how a developer could use its MCP server to provision an Azure Kubernetes Service (AKS) cluster. The developer issues natural-language instructions to an AI assistant, prompting the AI to execute MCP tools that invoke Pulumi CLI commands.
MCP caveats
Just as vibe coding isn’t a fit for every project, MCP isn’t the best option for every use case either. According to MCP experts, these servers can be unnecessary when they sidestep standard CLIs.
They can also introduce major security risks. This tracks with AI use in general, as 62% of IT leaders cite security and privacy risks as the top AI concern, according to the AI in DevOps report by Enterprise Management Associates (EMA).
As such, it’s best to test out these MCP servers with low-risk permissions, like read-only capabilities, before testing write functions. And use them only with trusted LLMs and trusted MCP clients.
Also, beware of exposing high-value, long-lived privileges to MCP clients. Because AI coding agents are based on nondeterministic LLMs, their behavior can be unpredictable. Throw in autonomous control over mutable devops functions, and you could land in all kinds of trouble, ranging from broken deployments to runaway token usage.
Lastly, using the official MCPs above, as opposed to community-supported libraries, will probably guarantee longer longevity and ongoing maintenance, too.
Early MCP success stories
Although it’s still early days with MCP and agents, there’s a sense of cautious optimism as proven MCP workflows emerge.
Take Block’s journey. Through company-wide use of its MCP-compatible agent, Goose, 12,000 employees are now utilizing agents and MCP for “increasingly creative and practical ways to remove bottlenecks and focus on higher-value work,” writes Angie Jones, head of developer relations.
Other engineers report using MCP servers to enhance workflows that are devops-adjacent, like the Filesystem MCP server for accessing local files, the Linear MCP server for issue tracking, the Chrome DevTools MCP server for browser debugging, and the Playwright MCP server for continuous testing.
And beyond the official MCP servers mentioned above, many community-supported MCPs are emerging for Docker, Kubernetes, and other cloud-native infrastructure utilities.
Devops comes with toil and cost. So, the case to level it up with MCP is strong. As long as you keep controls safe, it should be fun to see how these MCP servers integrate into your work and impact your productivity. Happy MCP-opsing.