Best MCP Servers for DevOps in 2026

The best MCP servers for DevOps in 2026 come from the clouds (AWS, Azure, Google Cloud), infrastructure-as-code tools (Terraform, Pulumi), CI/CD services (CircleCI, Buildkite) and app platforms (Render, Heroku, DigitalOcean). Each lets an AI assistant read build logs, inspect cloud resources or start a deploy, using permissions you already manage.

MCP (Model Context Protocol) is the open standard that connects AI apps such as Claude Code, Cursor, Codex and GitHub Copilot to outside tools. For DevOps that means asking "Why did last night's pipeline fail?" or "Which S3 buckets have versioning off?" and getting an answer from your real systems.

We checked every entry against the vendor's own site or docs on 28 October 2026. We did not test these products.

How we chose

We used the MCP Guide's standard listicle criteria:

• Official. The vendor publishes its own MCP server. Community projects, including the popular Kubernetes server, are mentioned only in the FAQ.
• Documented access control. The docs explain how the server signs in and what it can change.
• Verified today. We opened each vendor's docs and pricing page on 28 October 2026. Prices come only from official pricing pages.

This list focuses on running infrastructure. Code hosting, preview deploys and browser debugging (GitHub, GitLab, Vercel, Netlify, Cloudflare and others) are on our best MCP servers for developers list. Monitoring tools are on the best MCP servers for observability. The order below is not a ranking. More lists live on the MCP Guide's best-of hub.

Comparison at a glance

1. AWS: your whole account through IAM

What it is. The AWS MCP Server is part of the Agent Toolkit for AWS. It lets agents search AWS documentation, check Region availability and run AWS API calls, including Python scripts in a sandbox. It adds the aws:ViaAWSMCPService and aws:CalledViaAWSMCP condition keys to every request, "and CloudTrail records the calls."

Best for: Teams that run on AWS and want agents to work under existing IAM controls.

Pricing: AWS says you can use it "at no additional charge." You pay standard prices for resources the agent uses.

Pros:

• IAM condition keys let you write rules just for agent calls.
• OAuth 2.1 through AWS Sign-in, or SigV4 through a proxy.
• Documentation tools need no credentials.

Cons:

• The agent can do whatever its IAM role allows, so scope it down.
• OAuth sessions are bound to one IAM role and can't switch accounts.

Checked against: AWS docs on 28 October 2026. Full profile: AWS MCP server.

2. Azure: Azure resources in plain language

What it is. Microsoft says the Azure MCP Server "enables AI agents and clients to interact with Azure resources using natural language commands," such as listing storage accounts or running KQL queries. It installs through VS Code, Visual Studio, npm, NuGet, PyPI or Docker, and can be self-hosted as a remote endpoint on Azure Container Apps.

Best for: Microsoft shops working in Visual Studio, VS Code and GitHub Copilot.

Pricing: Microsoft lists no separate MCP price. The overview says you need "A GitHub Copilot subscription" and an Azure account to use the tools in supported hosts.

Pros:

• Entra ID sign-in and Azure RBAC.
• Built into Visual Studio 2022 version 17.14.30 and later.
• Open source on GitHub.

Cons:

• Microsoft says the local server "is intended strictly for developer use within your organization."
• Remote use means hosting it yourself.

Checked against: Microsoft Learn on 28 October 2026. Full profile: Azure MCP server.

3. Google Cloud: a remote endpoint per product

What it is. Google Cloud's remote MCP servers run on Google's infrastructure. Its supported products page lists endpoints for Cloud Run, Compute Engine, GKE, Cloud Logging and Cloud Monitoring, among others. Several other products, including the Google Cloud CLI server, are marked Preview.

Best for: Teams on GKE, Cloud Run or Compute Engine.

Pricing: Google's MCP overview lists no MCP price. The underlying services bill as usual.

Pros:

• Hosted by Google, one endpoint per product.
• Sign in with Google credentials or an app identity.
• Google documents how to prevent read-write tool use.

Cons:

• You connect product by product.
• Status varies, with many endpoints in Preview.

Checked against: Google Cloud docs on 28 October 2026. Full profile: Google Cloud MCP server.

4. Terraform: providers, modules and workspaces

What it is. The Terraform MCP Server gives AI models live access to provider docs, modules and policies, and to workspaces in HCP Terraform and Terraform Enterprise. HashiCorp hosts a remote version at https://mcp.terraform.io/mcp and publishes the code under MPL-2.0.

Best for: Platform teams writing Terraform who want current provider docs instead of stale model memory.

Pricing: The server is free and open source. HashiCorp's pricing page lists HCP Terraform Essentials "Starting at $0.10 per month per resource," with Standard and Premium tiers above it.

Pros:

• Provider docs that match today's versions.
• Tools needing approval are off unless you set ENABLE_TF_OPERATIONS.
• Hosted and self-hosted options.

Cons:

• Workspace tools can create, update and delete workspaces.
• Needs an HCP Terraform or Terraform Enterprise token for workspace work.

Checked against: HashiCorp's GitHub repository on 28 October 2026. Full profile: Terraform MCP server.

5. Pulumi: stacks, policies and Neo

What it is. Pulumi hosts its MCP server at https://mcp.ai.pulumi.com/mcp. Tools cover stacks, policy violations, resource search, the Pulumi Registry and code generation. Pulumi says "the server itself never runs pulumi up or changes your infrastructure."

Best for: Pulumi users who want an agent to look up stacks and policy issues.

Pricing: Pulumi's pricing page lists Free, Essentials at $40 a month, Pro at $400 a month and Enterprise at $2,000 a month. Registry tools are free, resource-search needs a paid edition, and "Neo usage is billed to your organization."

Pros:

• Mostly read-only tools.
• Every call runs as the token's identity.
• Headless mode for CI with a bearer token.

Cons:

• The neo-bridge tool can lead to infrastructure changes once a Neo request is approved.
• Some tools need a paid edition.

Checked against: Pulumi docs on 28 October 2026. Full profile: Pulumi MCP server.

6. CircleCI: pipelines, logs and reruns

What it is. CircleCI's hosted MCP server at https://mcp.circleci.com/v1/mcp exposes 13 tools, such as get_job_logs, list_job_tests, rerun_workflow and cancel_workflow. It authenticates with "OAuth2 or a personal API token." CircleCI says its older local server "is deprecated."

Best for: Teams who want an AI to read failing build logs and rerun workflows.

Pricing: No separate MCP price. CircleCI's pricing page lists a Free plan with 30,000 credits a month and Performance starting at $15 a month.

Pros:

• Hosted, with nothing to install.
• OAuth sign-in.
• A CLI-based MCP option for local editors.

Cons:

• If you set up the old npm server, you need to switch.
• Rerun and cancel tools act on real pipelines.

Checked against: CircleCI docs on 28 October 2026. Full profile: CircleCI MCP server.

7. Buildkite: CI with a read-only endpoint

What it is. The Buildkite MCP server has a standard OAuth endpoint, a read-only endpoint at https://mcp.buildkite.com/mcp/readonly and a token pass-through endpoint for headless agents. A local server is also available. Buildkite says it "is under active development."

Best for: Teams on Buildkite who want AI help with failed jobs and test results.

Pricing: No separate MCP price. Buildkite's pricing page lists a $0 Free plan and Pro at "$ 30 per active user/month."

Pros:

• A dedicated read-only endpoint.
• Short-lived OAuth tokens tied to the user.
• Hosted and local options.

Cons:

• No stability label yet.
• The standard endpoint can retry jobs and change state.

Checked against: Buildkite docs on 28 October 2026. Full profile: Buildkite MCP server.

8. Render: deploys, logs and metrics

What it is. The Render MCP server can create services and databases, trigger deploys, read logs and metrics, update environment variables and "Run a read-only SQL query" on Render Postgres. It does not delete resources or change scaling settings.

Best for: Small teams hosting apps on Render.

Pricing: No separate MCP price. Render's pricing page lists Hobby at $0, Pro at $25 a month and Scale at $499 a month, each plus compute.

Pros:

• No delete tools.
• OAuth through official plugins for Claude Code, Codex and Cursor.
• Read-only database queries.

Cons:

• Can still create billable services.
• Can't change scaling or most service settings.

Checked against: Render docs on 28 October 2026. Full profile: Render MCP server.

9. Heroku: apps, dynos and Postgres

What it is. The Heroku MCP Server runs locally through the Heroku CLI. Tools cover apps, deploys, dyno scaling, add-ons, pipelines, logs and Heroku Postgres. Heroku says it "is currently in early development."

Best for: Heroku users who live in the CLI.

Pricing: No separate MCP price. Heroku's pricing page lists Eco dynos at $5 a month and Basic at $7.

Pros:

• Uses your existing Heroku CLI session.
• Covers Postgres diagnostics.
• Pipeline promotion from chat.

Cons:

• Tools such as pg_kill and ps_scale act on production.
• Local only, with no hosted endpoint.

Checked against: Heroku Dev Center on 28 October 2026. Full profile: Heroku MCP server.

10. DigitalOcean: 25 remote services and a router

What it is. DigitalOcean hosts remote MCP servers for 25 services, including Droplets, Kubernetes, App Platform, Databases and Spaces. A single MCP Router at https://mcp.digitalocean.com/mcp covers all but one of them.

Best for: Startups and small teams running on DigitalOcean.

Pricing: No separate MCP price. DigitalOcean's Droplet pricing page bills compute by the hour.

Pros:

• DigitalOcean recommends OAuth "instead of a static API token."
• Connect one service or use the router.
• Code is open source, per our profile.

Cons:

• A token with full scope can manage everything.
• The router exposes many tools at once.

Checked against: DigitalOcean docs on 28 October 2026. Full profile: DigitalOcean MCP server.

Guardrails for agents with infrastructure access

A DevOps agent can do real damage in seconds. Treat it like a new engineer on their first day.

1. Give it its own identity. Use a dedicated IAM role, service principal or token, not your admin login.
2. Start read-only. Use Buildkite's read-only endpoint, a read-only IAM policy or Google's guidance on blocking read-write tools.
3. Gate changes. Keep Terraform's ENABLE_TF_OPERATIONS off and deny Pulumi's neo-bridge until you trust the workflow.
4. Separate environments. Point agents at staging first. Production access should be a deliberate step.
5. Log everything. CloudTrail and the AWS condition keys show which calls came from an agent. See MCP excessive permissions and the MCP security checklist.

Next step

If you want agents that can work across your cloud and CI without broad admin keys, our MCP development team can design the roles, gateways and approvals. To find the gaps first, book an MCP readiness audit.

Frequently asked questions