Skip to main content
Comparison10 min read·Updated September 7, 2026
🧩

Terraform MCP vs AWS MCP vs gcloud MCP (2026 Guide)

B

A. Frans

Published September 7, 2026

MCP ServersTerraformAWSGoogle CloudDevOps

If you are about to hand an AI agent your cloud credentials, the interesting question is not which MCP server ships more tools. It is what happens when the agent is confidently wrong at 2am.

That question splits these three apart faster than any feature list. HashiCorp's Terraform server, Google's gcloud server, and the community AWS server all get filed under "cloud MCP" in every directory including ours, and they occupy completely different positions on the risk curve. One of them mostly reads documentation. One of them runs CLI commands against your live project. One of them is a third party you have never met holding a shell.

The three at a glance

HashiCorp Terraformgcloud MCPAWS MCP Server
MaintainerHashiCorpGoogle (googleapis)alexei-led, community
Trust tier on this siteOfficialOfficialCommunity
Repohashicorp/terraform-mcp-servergoogleapis/gcloud-mcpalexei-led/aws-mcp-server
LicenseMPL-2.0Apache-2.0MIT
What it touchesTerraform Registry: providers, modules, docsYour GCP project through the gcloud CLIYour AWS account through the AWS CLI
Can it change infrastructure?Not directlyYes, if you let itYes, if you let it
Runs commands in a sandboxN/ANoYes, containerized
Best first useWriting correct TerraformAsking what is running and whyRead-only account exploration

What each one wires up

HashiCorp Terraform: the one that makes the agent stop inventing arguments

This is the least dangerous and the most immediately useful of the three, which is an unusual combination.

The problem it solves is narrow and painfully familiar. Ask any model to write Terraform and it will produce something that looks right, uses a resource argument that was removed two provider versions ago, and fails on plan. The model is recalling the shape of Terraform from training data instead of reading the provider documentation for the version you have pinned.

HashiCorp's server puts the Terraform Registry in the loop. The agent can look up a provider, check a module, and read the actual argument reference before it writes the block. Correctness goes up sharply and you have granted almost nothing, because registry lookups are public reads.

If you write any Terraform at all, add this one first. It is the cheapest correctness win available.

For the authoring conventions on top of it, Terraform Skill by antonbabenko covers module structure, testing, and CI patterns for Terraform and OpenTofu.

gcloud MCP: Google's own wrapper around the CLI you already use

Google's server gives the agent the gcloud CLI. That is the whole idea, and it is a good one, because gcloud already has your auth, your project context, and thirteen years of surface area.

What it is good at is the question you ask three times a week and hate looking up: which service accounts have this role, why is this Cloud Run revision not receiving traffic, what changed in this project yesterday. The agent runs the command, reads the output, and runs the follow-up without you switching windows.

The catch is that gcloud is not a read-only tool. The same interface that answers "what is running" can also delete it. Scope the credentials before you scope the ambition. Give the agent a service account with viewer roles and add write permissions deliberately, one at a time, when a specific workflow needs them.

If your deployment target is Cloud Run specifically, cloud-run-mcp from GoogleCloudPlatform is a narrower server built for that one job, and narrower is usually safer.

AWS MCP Server: a shell, in a box, from someone you do not know

The community AWS server executes AWS CLI commands, and its main design decision is that it runs them inside a container rather than directly on your machine. That is a real mitigation and worth crediting.

It is also the one that deserves the most scrutiny before installation, for a reason that has nothing to do with code quality. AWS publishes no equivalent first-party server in this directory, so the AWS gap gets filled by community projects. When you install one, you are trusting a maintainer you have not met with an interface to your production account.

That is not a reason to avoid it. It is a reason to read the source before you run it, pin a commit rather than tracking main, and give it an IAM role scoped to exactly what you need. Do those three things and it becomes a useful way to explore an AWS account you inherited and do not understand.

Alternatives in the same space if you want to compare: aws-mcp by RafalWilinski and aws-skills by zxkane, which leans toward CDK, SST, and serverless patterns rather than raw CLI access.

Official versus community matters more here than usual

For most agent skills, the trust tier is a soft signal. A community Markdown formatter that goes wrong wastes ten minutes.

Cloud servers are different because the failure mode has a bill attached. A compromised or careless server with write credentials can spin up resources, exfiltrate configuration, or delete something that was not backed up. The blast radius is your infrastructure.

So the ordering is not snobbery. Prefer official when an official option exists, read the source when it does not, and never skip the source review because a repo has a lot of stars. Stars measure popularity, not review.

One honest caveat about the official label: Microsoft's Azure DevOps MCP is official and frequently listed alongside these three, but it covers boards, repos, and pipelines rather than Azure resource management. If you came looking for an Azure equivalent of gcloud MCP, that is not what this is. Check the scope, not the logo.

The permission question, concretely

Every one of these servers inherits whatever credentials sit in your environment. There is no separate permission model. If your terminal can delete the production database, so can the agent using your terminal.

Three habits that cost almost nothing:

  • Create a dedicated identity. A service account or IAM role used only by the agent, starting at read-only. When you need a write action, add that specific permission and note why.
  • Separate profiles by environment. The agent should not have a credential that can reach production while you are working on staging. Profile switching is friction worth keeping.
  • Assume the agent will retry. Agents retry failed commands. A command that half-succeeded and then retried is how you end up with two of something expensive. Read-only defaults make this a non-event.

Cost visibility, which nobody sets up until it hurts

Worth adding alongside whichever server you pick: aws-finops-mcp-server by ravikiranvm analyses spend, audits for waste, and pulls budget data into the same conversation where you are making the change. AWS Cost Saver covers similar ground with a focus on unused resources and right-sizing.

Both are read-only in nature, which makes them the easiest thing on this page to justify installing. The agent that can see the bill gives noticeably better infrastructure advice than the one that cannot, and asking "what would this change cost per month" before applying it is a habit worth building.

Installing without giving away the keys

A note on install commands: skip the ones you find in directories, including the one on this site. Every MCP server publishes its exact invocation in its own README, those commands change with releases, and a stale copy is how people end up debugging a package name for twenty minutes.

The shape is stable even when the specifics are not. In Claude Code:

claude mcp add <name> -- <command from the project README>

Then check what actually loaded before you trust it:

claude mcp list

Sequence that avoids surprises: read the README, read enough of the source to know what it executes, provision a read-only credential, install, ask the agent a question you already know the answer to, and only then let it near anything that matters.

Which to add first

Terraform MCP, today, regardless of cloud. Public registry reads, no credentials, immediate improvement in generated configuration. There is no argument against it.

gcloud MCP if you are on GCP, with a viewer service account, because it is first-party and the CLI it wraps is already the tool you use.

AWS MCP Server if you are on AWS, after reading the source, pinned to a commit, with a scoped IAM role. The container sandbox is a real advantage over servers that run commands directly on your host.

What I would not do is install all three on day one to see what happens. Each one changes what your agent can reach. Add them one at a time, with a week between, so you can attribute both the wins and the surprises.

More on the surrounding workflow in our infrastructure-as-code skills roundup and the cloud cost optimisation guide. If the MCP-versus-skill distinction is still fuzzy, we covered that separately. Broader tooling by role sits on our developers page.

FAQ

Can the Terraform MCP server run terraform apply for me?

That is not what it is for. It brings registry knowledge, providers, modules, and documentation into the agent's context so the configuration it writes is correct for your provider version. Running plan and apply stays in your terminal or your pipeline, which is the right place for it.

Is there an official AWS MCP server?

Not one listed in this directory, which is why the AWS entries here are community-maintained. Treat that as a reason for source review and scoped credentials rather than a reason to skip AWS tooling entirely.

Can I run more than one cloud MCP server at once?

Yes, and multi-cloud teams often do. The practical cost is context: every server adds tool definitions to the agent's working context, and a crowded toolset makes tool selection worse. Add them one at a time and remove the ones you have not used in a month.

Does the containerized sandbox in the AWS server make it safe?

It contains the process, not the credentials. A command that deletes an S3 bucket deletes it whether it ran in a container or not. The sandbox limits what the server can do to your laptop. IAM is what limits what it can do to your account.

What is the minimum safe setup for trying any of these?

A separate credential with read-only permissions, a non-production project or account, and a first question whose answer you already know. If the agent gets that one wrong, you have learned something cheaply.

Share this article

📬

Get More AI Tool Guides

New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.