<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Ankit Verma</title>
    <description>The latest articles on DEV Community by Ankit Verma (@dockndevai).</description>
    <link>https://dev.to/dockndevai</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4090691%2F3aebacd1-e265-44ab-b11a-572397751831.png</url>
      <title>DEV Community: Ankit Verma</title>
      <link>https://dev.to/dockndevai</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/dockndevai"/>
    <language>en</language>
    <item>
      <title>8 safe-by-default MCP servers for infra — and the governance model behind them</title>
      <dc:creator>Ankit Verma</dc:creator>
      <pubDate>Sun, 23 Aug 2026 16:26:47 +0000</pubDate>
      <link>https://dev.to/dockndevai/8-safe-by-default-mcp-servers-formcp-ai-devops-opensource-infra-and-the-governance-model-4gi8</link>
      <guid>https://dev.to/dockndevai/8-safe-by-default-mcp-servers-formcp-ai-devops-opensource-infra-and-the-governance-model-4gi8</guid>
      <description>&lt;p&gt;Giving an AI agent access to production infrastructure is a great way to move fast — and a great way to have it drop a database because a prompt was ambiguous. I wanted the upside without the footguns, so I built a family of &lt;a href="https://modelcontextprotocol.io" rel="noopener noreferrer"&gt;Model Context Protocol&lt;/a&gt; servers that share one safe-by-default governance model.&lt;/p&gt;

&lt;h2&gt;
  
  
  The problem
&lt;/h2&gt;

&lt;p&gt;MCP lets an agent call tools. For infra, those tools can be &lt;code&gt;delete_topic&lt;/code&gt;, &lt;code&gt;DROP TABLE&lt;/code&gt;, &lt;code&gt;delete_resource_group&lt;/code&gt;. The usual answer is "just don't expose the dangerous ones" — but then the server isn't useful when you &lt;em&gt;do&lt;/em&gt; need them. I wanted graduated, explicit control instead.&lt;/p&gt;

&lt;h2&gt;
  
  
  The governance model
&lt;/h2&gt;

&lt;p&gt;Every server shares the same layered policy engine:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Access modes&lt;/strong&gt; — &lt;code&gt;read-only&lt;/code&gt; → &lt;code&gt;read-write&lt;/code&gt; → &lt;code&gt;admin&lt;/code&gt;. Tools above the current mode are &lt;em&gt;never even registered&lt;/em&gt;, so the model can't call what it can't see.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Allowlists&lt;/strong&gt; — scope to specific realms / namespaces / topics / databases / projects / subscriptions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Protected resources&lt;/strong&gt; — system/prod resources (kube-system, the &lt;code&gt;system&lt;/code&gt; DB, internal Kafka topics, the master realm) are readable but never mutable.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Destructive gating&lt;/strong&gt; — deletes need an explicit &lt;code&gt;ALLOW_DELETE&lt;/code&gt; flag on top of admin mode.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Typed confirmation&lt;/strong&gt; — high-impact ops (delete a project / resource group) require a &lt;code&gt;confirm&lt;/code&gt; value that echoes the exact target name. A boolean isn't enough.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Secret redaction&lt;/strong&gt; — credentials are stripped before anything reaches the model.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dry-run + audit&lt;/strong&gt; — preview writes without executing; every guarded op emits a JSON audit line to stderr.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The servers
&lt;/h2&gt;

&lt;p&gt;All MIT-licensed, TypeScript, published on npm as &lt;code&gt;@dockndevai/mcp-*&lt;/code&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;mcp-kubernetes&lt;/strong&gt; — pods, logs, deployments, scale/restart, apply, exec&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;mcp-kafka&lt;/strong&gt; — topics, consumer groups + lag, create/alter/reset&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;mcp-clickhouse&lt;/strong&gt; — schema, queries, SQL-classified read/write/destructive gating&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;mcp-debezium&lt;/strong&gt; — CDC connector status, config, lifecycle&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;mcp-oci&lt;/strong&gt; — Oracle Cloud discovery + Terraform generation&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;mcp-azure&lt;/strong&gt; — Azure Resource Manager inventory, tags, VM power, lifecycle&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;mcp-azure-devops&lt;/strong&gt; — boards, repos, pipelines, projects&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;mcp-keycloak&lt;/strong&gt; — realms, users, clients, roles, groups&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try one
&lt;/h2&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;npx -y @dockndevai/mcp-kubernetes
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Or in Claude Code:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;claude mcp add kubernetes -e KUBECONFIG_PATH=~/.kube/config -e K8S_MODE=read-only -- npx -y @dockndevai/mcp-kubernetes
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Every repo has per-client setup for Claude, Cursor, Codex, VS Code, and Windsurf.&lt;/p&gt;

&lt;p&gt;Repos: &lt;a href="https://github.com/dockndevai" rel="noopener noreferrer"&gt;https://github.com/dockndevai&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I'd love feedback — especially on whether the mode + typed-confirmation split is the right default for infrastructure MCP servers.&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>ai</category>
      <category>devops</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
