MCP server de cloud-tools

Multi-cloud cost, inventory and waste analysis in Rust — exposed as an MCP server so an AI agent can query spend and waste directly, and as an HTTP API for everything else.

It reports spend and waste: idle instances, oversized nodes, previous-generation hardware, unattached resources, and commitments you are not using.

Eight tools, one shape

Every tool takes cloud and nothing else it does not need. Credentials never appear in a tool call — the server reads them from its own environment, so the agent never handles a secret and never needs to be told one.

{ "cloud": "aws" }
{ "cloud": "gcp" }

Start with check_access. It takes no arguments, contacts no cloud, and reports which clouds this server can reach and what is missing for the ones it cannot.

Tool	AWS	GCP	Cloudflare	OVH
check_access	reports configuration for all four
get_costs	yes	needs a billing table	yes	yes
compare_costs	yes	needs a billing table	—	—
get_inventory	—	yes	yes	yes
get_waste	yes	yes	—	—
get_commitments	yes	yes	—	—
get_recommendations	—	yes	—	—
get_cross_cloud_summary	every cloud the server can reach

A dash is not a silent empty result. The tool answers, for example, waste analysis is not implemented for ovh; supported: aws, gcp, so an agent cannot read a gap as a clean bill.

Reading a zero correctly

get_waste and get_inventory return a coverage field and an errors list. A zero total means one of two entirely different things:

coverage	meaning
complete	Nothing is wasted.
PARTIAL — N API call(s) failed…	Part of the account could not be read.

A partial result is not a clean bill of health.

Configuring the server

Credentials are read from the environment of the machine running cloud-tools. Nothing is stored, and every call is read-only.

AWS

Resolved in this order, first match winning:

1. AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY
2. ~/.aws/credentials, profile from AWS_PROFILE, otherwise [default]
3. An ECS task role
4. An EC2 instance role

An SSO-only profile is not read directly. Export it first: aws configure export-credentials --profile <name> --format env.

To scan a different account, pass a role to assume. A role ARN is not a secret, so it stays an argument:

{ "cloud": "aws", "target": { "role_arn": "arn:aws:iam::…:role/ReadOnly" } }

CLOUD_TOOLS_AWS_ROLE_ARN and CLOUD_TOOLS_AWS_EXTERNAL_ID set a default for that, if every call should assume the same role.

get_waste reads EC2 instances, volumes, snapshots, AMIs, Elastic IPs, key pairs and reserved instances; RDS; S3; Lambda; DynamoDB; ECS; ElastiCache; load balancers; NAT gateways; CloudWatch log groups with no retention; and Compute Optimizer. Idle and oversized come from CloudWatch CPU series, so they are measured rather than guessed.

GCP

Application Default Credentials — gcloud auth application-default login is enough — or GOOGLE_APPLICATION_CREDENTIALS pointing at a service-account file.

Variable	Purpose
CLOUD_TOOLS_GCP_PROJECTS	Comma-separated project IDs to query
CLOUD_TOOLS_GCP_BILLING_TABLE	BigQuery billing export, project.dataset.table

Projects can also be chosen per call with target.project_ids, which is how one agent scans several.

billing_table matters. Google publishes no per-service spend API, so real costs come from the BigQuery billing export. Without it, get_costs falls back to the Budgets API and labels the result "source": "budgets" with a note saying these are budget amounts, not spend. compare_costs requires it outright.

Cloudflare

CLOUDFLARE_API_TOKEN     a token with read access to account resources
CLOUDFLARE_ACCOUNT_ID

OVH

OVH_APPLICATION_KEY
OVH_APPLICATION_SECRET
OVH_CONSUMER_KEY
OVH_ENDPOINT             ovh-eu (default), ovh-us or ovh-ca

Install

Not released yet. The newest tag is v0.1.0, which shipped .tar.gz archives under the old asset names, and @munhq/cloud-tools is not on npm. The three commands below start working the moment v0.2.0 is tagged. Until then, build from source: cargo build --release.

As an MCP server

claude mcp add cloud-tools -- npx -y @munhq/cloud-tools

Anything that reads a JSON config — Claude Desktop, Cursor, Windsurf, Zed, Cline:

{ "mcpServers": { "cloud-tools": { "command": "npx", "args": ["-y", "@munhq/cloud-tools"] } } }

The npm package is a small wrapper, because the server is a compiled binary and not JavaScript. On install it resolves the release asset for your platform, verifies it against the published SHA256SUMS, caches it under ~/.cache/cloud-tools/bin/ and executes it.

As a binary

Prebuilt binaries for Linux, macOS and Windows, x86_64 and arm64, are attached to each release. The script below downloads the one for your machine and checks it against SHA256SUMS:

curl -sSL https://raw.githubusercontent.com/munhq/cloud-tools/main/install.sh | bash

Or take a single asset directly:

curl -fsSLO https://github.com/munhq/cloud-tools/releases/latest/download/cloud-tools-x86_64-linux
chmod +x cloud-tools-x86_64-linux

As a container

docker run -i --rm \
  -e AWS_ACCESS_KEY_ID -e AWS_SECRET_ACCESS_KEY -e AWS_REGION \
  munhq/cloud-tools

Running it

Built as a library (cloud_tools) with an optional binary. Two feature flags, both on by default:

• mcp — MCP server over stdio and streamable HTTP, via rmcp
• http — HTTP API via axum

cargo build --release
./target/release/cloud-tools            # MCP server on stdio

As an MCP server it plugs into any MCP-capable agent. The agent calls the tools above directly and reasons over the results.

To embed it instead, depend on the library and turn the binary features off:

cloud-tools = { git = "https://github.com/munhq/cloud-tools", default-features = false }

Status

Working and used against live AWS, GCP, Cloudflare and OVH accounts. Waste analysis exists for AWS and GCP only — Cloudflare and OVH have none at all, not merely fewer rules, and get_waste says so rather than returning an empty list. AWS has no inventory tool yet, though the listing calls behind one already exist. Contributions adding provider coverage are welcome.

Privacy Policy

cloud-tools collects nothing. There is no telemetry, no analytics and no phone-home of any kind — verified by the fact that the only hosts in the source are cloud provider APIs plus github.com, which the npm wrapper uses to download the binary for your platform.

What is collected. Nothing. No account, no sign-up, no identifier.

What is processed, and where. Credentials you pass are used to sign requests to your own cloud provider and are held in memory for the life of that call. The server runs on your machine, so nothing is sent to munhq or to any third party. The responses go back to the MCP client that asked for them and nowhere else.

Storage and retention. Nothing is written to disk and nothing is retained. The one exception is the npm wrapper, which caches the downloaded binary under ~/.cache/cloud-tools/bin/ so it is not re-fetched on every run. That cache holds a program, never your data or your credentials.

Logging. Diagnostics go to stderr, which your MCP client captures. Credentials are never logged. Error messages quote the provider's own response, which may name a project, an account id or a resource — the same information the tool was asked to report.

Third-party sharing. None. The only network destinations are the cloud APIs of the providers whose credentials you supply: AWS (*.amazonaws.com), Google Cloud (*.googleapis.com), Cloudflare (api.cloudflare.com) and OVH (*.api.ovh.com, api.us.ovhcloud.com).

Permissions. Read-only is sufficient for every tool. All eight declare readOnlyHint: true, and nothing in this server creates, modifies or deletes a cloud resource.

Contact. hello@munhq.com. The canonical version of this policy is at https://munhq.com/privacy/cloud-tools.

Licence

Apache-2.0. See LICENSE.