About cloud-tools
cloud-tools is a Rust MCP server and HTTP API for querying cloud spend, inventory, waste, and commitments across AWS, GCP, Cloudflare, and OVH. It uses credentials from its runtime environment and performs read-only queries.
Key features
- 01Cost and cross-cloud spend queries
- 02Resource inventory where supported
- 03Waste analysis for AWS and GCP
- 04Commitment and recommendation reports
- 05Access and coverage status
Use cases
- 01Review cloud spend across providers
- 02Identify idle or oversized resources
- 03Check provider access and incomplete results
- 04Inspect commitments and GCP recommendations
MCP support
Exposes eight read-only tools over MCP stdio or streamable HTTP for agents to query supported cloud providers. Credentials remain in the server environment rather than tool calls.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @munhq/cloud-toolsCLOUD_TOOLS_MODEmcp for an MCP client over stdio, http for the REST API. Defaults to mcp, which is what every MCP client wants.Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEMulti-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:
AWS_ACCESS_KEY_IDandAWS_SECRET_ACCESS_KEY~/.aws/credentials, profile fromAWS_PROFILE, otherwise[default]- An ECS task role
- 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.gzarchives under the old asset names, and@munhq/cloud-toolsis not on npm. The three commands below start working the momentv0.2.0is 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, viarmcphttp— 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. [email protected]. The canonical version of this policy is at https://munhq.com/privacy/cloud-tools.
Licence
Apache-2.0. See LICENSE.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.