GCP Platform Engineer

platform-mcp

mcp-name: io.github.deBilla/platform-mcp

A read-only Model Context Protocol server that turns an AI agent (Claude Code, Claude Desktop, or any MCP client) into a GCP platform engineer. Point it at your Google Cloud projects and ask it to investigate incidents, take inventory, and surface cost-optimization opportunities — all without any ability to change your infrastructure.

Observation only. No tool in this server mutates state. Combined with a viewer-only identity (below), that gives you a hard, defense-in-depth guarantee that an agent can look but never touch.

What it can do

Typical prompts once it's connected:

• "What are the top error groups in the last 24 hours, and which one is newest?"
• "Which GKE node pools are over-provisioned? Show mean CPU against machine type."
• "Where can I reduce spend in this project?"

Multiple environments

One server can reach several projects. Define them under PLATFORM_MCP_ENVIRONMENTS (see Configuration) and the agent picks one from the wording of your prompt:

• "Any errors in staging in the last hour?"
• "Compare Cloud Run services between staging and prod."

Every tool takes an optional environment argument. Omit it and the default environment is used; pass environment="production" to target another. Names, any aliases you define, common shorthands (prod, stg, qa, …) and bare project ids all resolve. An unrecognized name is an error listing the valid options — a typo can never silently retarget the wrong project.

Each environment carries its own service account, so staging and production are reached through separate identities from the same process, and every result echoes back the environment and project it came from.

Requirements

• Python 3.11+
• A Google Cloud project and credentials (your own login, or a service account)
• The gcloud CLI for the one-time setup

Install

uvx platform-mcp          # no install step; uv fetches it on demand
pipx install platform-mcp # or keep it on PATH

From a checkout, for development:

git clone https://github.com/deBilla/platform-mcp.git
cd platform-mcp
python3 -m venv .venv
./.venv/bin/pip install -e ".[dev]"

Check your setup

platform-mcp doctor

This checks, for every configured environment, that Application Default Credentials exist, that the read-only service account can be impersonated, that a real API read succeeds, and that the billing export is readable — printing the exact command to fix whatever fails. Run it before reporting a problem.

One-time GCP setup

Run these once per project you want to reach — staging and production each need their own APIs enabled and their own read-only service account.

1. Enable the APIs the tools depend on:

gcloud services enable \
  logging.googleapis.com monitoring.googleapis.com clouderrorreporting.googleapis.com \
  recommender.googleapis.com cloudasset.googleapis.com cloudbilling.googleapis.com \
  bigquery.googleapis.com \
  --project YOUR_PROJECT_ID

2. Grant read-only access to the identity the server runs as.

For local development with your own login (Application Default Credentials):

gcloud auth application-default login

The identity needs these viewer roles on the project, plus roles/billing.viewer on the billing account:

roles/viewer                # broad read (compute, run, gke, sql via Asset Inventory)
roles/logging.viewer
roles/monitoring.viewer
roles/errorreporting.viewer
roles/recommender.viewer
roles/cloudasset.viewer
roles/bigquery.dataViewer    # only for get_cost_breakdown
roles/bigquery.jobUser       # only for get_cost_breakdown

3. (Recommended) Use a dedicated read-only service account instead of your login. platform-mcp setup does every step below, is safe to re-run, and prints the config stanza at the end. It ships with the package, so there is nothing to clone:

uvx platform-mcp setup \
  --project YOUR_PROJECT_ID \
  --user [email protected] \
  --billing-dataset YOUR_BILLING_PROJECT:billing   # optional

Or by hand:

PROJECT=YOUR_PROJECT_ID
gcloud iam service-accounts create platform-mcp-ro \
  --display-name "platform-mcp read-only" --project $PROJECT

SA=platform-mcp-ro@$PROJECT.iam.gserviceaccount.com
for ROLE in roles/viewer roles/logging.viewer roles/monitoring.viewer \
  roles/errorreporting.viewer roles/recommender.viewer roles/cloudasset.viewer \
  roles/bigquery.jobUser; do
  gcloud projects add-iam-policy-binding $PROJECT \
    --member="serviceAccount:$SA" --role="$ROLE" --condition=None
done

# Let your own login impersonate it (no key file to manage):
gcloud iam service-accounts add-iam-policy-binding $SA \
  --member="user:[email protected]" \
  --role="roles/iam.serviceAccountTokenCreator" --project $PROJECT

The grant everyone forgets. roles/bigquery.jobUser above only lets the account start a query; it grants no access to any data. A billing export almost always lives in a different project, so the account also needs read on that dataset. Without it get_cost_breakdown returns 403 while every other tool works, which reads like a bug in the tool rather than a missing grant:

bq add-iam-policy-binding \
  --member="serviceAccount:$SA" --role=roles/bigquery.dataViewer \
  YOUR_BILLING_PROJECT:billing

If you lack admin on the billing project, that one line is what to send to someone who has it. platform-mcp doctor checks it and says which side is missing.

Then reference it as that environment's impersonate value in PLATFORM_MCP_ENVIRONMENTS (preferred — no key file), or point at a downloaded key via GOOGLE_APPLICATION_CREDENTIALS.

Impersonation is performed by whatever identity your ADC resolves to. If your ADC is itself an impersonated service account, that SA — not your user — needs roles/iam.serviceAccountTokenCreator on each platform-mcp-ro.

Security model

Read-only is enforced by IAM, not by OAuth scope. The server requests the broad cloud-platform scope and stays read-only purely because it never calls a mutating API. Do not rely on the code alone — run it under a viewer-only identity (step 3 above) so the credential itself is incapable of writing, regardless of what code executes. This gives you two independent layers: the server doesn't try to write, and the identity couldn't if it did.

With multiple environments this stays per-project: each environment authenticates as its own service account, so a staging identity is never used to reach production. Grant each one viewer-only access to its project alone.

Configuration

The friendliest option is a config file, which keeps project ids and service account emails out of every client config you own:

mkdir -p ~/.config/platform-mcp
cp config.toml.example ~/.config/platform-mcp/config.toml
$EDITOR ~/.config/platform-mcp/config.toml

With that in place, registering the server takes no environment variables at all. Point PLATFORM_MCP_CONFIG elsewhere to use a different file — a copy committed to your infrastructure repo, for instance.

Environment variables still work and always win over the file, so an existing setup keeps running unchanged and a one-off override needs no edit:

PLATFORM_MCP_ENVIRONMENTS holds a JSON object; each entry accepts:

A bare string value is shorthand for {"project": "..."}. As JSON inside .mcp.json the quotes must be escaped; unescaped it reads:

{
  "staging": {
    "project": "my-app-staging",
    "impersonate": "[email protected]"
  },
  "production": {
    "project": "my-app",
    "impersonate": "[email protected]",
    "billing_export_table": "my-app.billing.gcp_billing_export_v1_XXXXXX"
  }
}

Single-environment mode. If PLATFORM_MCP_ENVIRONMENTS is unset the server behaves as before, exposing one environment named default:

Register with a client

Claude Code — with a config file in place, this is the whole thing:

claude mcp add platform-mcp --scope user -- uvx platform-mcp

Claude Desktop — needs an absolute path to uvx; see Claude Desktop for the full walkthrough.

Without a config file, add the environment variables from .mcp.json.example to either form.

Skip the approval prompt

Every tool here is read-only, so approving each call individually adds nothing. Allow the whole server once, in Claude Code settings:

{ "permissions": { "allow": ["mcp__platform-mcp__*"] } }

The glob must sit after a literal mcp__<server>__ prefix — an unanchored pattern like mcp__* is ignored with a warning and approves nothing.

MCP Inspector — for interactive testing:

uvx --with 'mcp[cli]' mcp dev src/platform_mcp/server.py

Claude Desktop

Installing for a teammate who uses Claude Desktop rather than the CLI. It works the same as Claude Code, with one failure mode the CLI does not have.

Claude Desktop does not inherit your shell PATH. It launches from the Finder, so uvx and anything installed by Homebrew or uv are invisible to it. A config that says "command": "uvx" fails with ENOENT — the server never starts, and the error names the command rather than the reason. The command path must be absolute.

What does not break is credentials. Application Default Credentials are a file the Google libraries read directly, so gcloud is needed once, in a terminal, to create that file — not at runtime. The config file is found the same way, from the home directory, whatever launched the server.

Before you start: the admin grant

Each person impersonates the read-only service accounts, so an admin runs this once per person, per environment (the platform-mcp setup script does it for whoever runs it, but not for anyone else):

gcloud iam service-accounts add-iam-policy-binding \
  platform-mcp-ro@YOUR_PROJECT_ID.iam.gserviceaccount.com \
  --member="user:[email protected]" \
  --role="roles/iam.serviceAccountTokenCreator" --project YOUR_PROJECT_ID

Without it everything installs fine and doctor fails at the impersonation step.

1. Install uv (macOS)

An analyst's laptop usually has neither Homebrew nor the Xcode Command Line Tools, and nothing here needs them. Avoid git, make and the stock /usr/bin/python3 along the way — they are stubs that pop a dialog offering to install a gigabyte of developer tooling. uv is a standalone binary:

curl -LsSf https://astral.sh/uv/install.sh | sh
which uvx      # note this absolute path — Claude Desktop needs it

Typically /Users/<you>/.local/bin/uvx.

2. Install the Google Cloud CLI

The tarball bundles its own Python, so it needs nothing else. Avoid brew install --cask google-cloud-sdk: Homebrew itself requires the Command Line Tools. Pick the build for your chip — uname -m prints arm64 for Apple Silicon, x86_64 for Intel:

# Apple Silicon
curl -O https://dl.google.com/dl/cloudsdk/channels/rapid/downloads/google-cloud-cli-darwin-arm.tar.gz
tar -xzf google-cloud-cli-darwin-arm.tar.gz

# Intel — same, with the other file
# curl -O https://dl.google.com/dl/cloudsdk/channels/rapid/downloads/google-cloud-cli-darwin-x86_64.tar.gz
# tar -xzf google-cloud-cli-darwin-x86_64.tar.gz

./google-cloud-sdk/install.sh --quiet

3. Authenticate

./google-cloud-sdk/bin/gcloud auth application-default login

Log in as yourself. If your ADC is itself an impersonated service account, that account — not you — becomes the identity doing the impersonation, and it will not hold the grant above.

4. Write the config file

mkdir -p ~/.config/platform-mcp
touch ~/.config/platform-mcp/config.toml
open -e ~/.config/platform-mcp/config.toml   # or any editor

default_environment = "staging"

[environments.staging]
project = "my-app-staging"
impersonate = "[email protected]"

[environments.production]
project = "my-app"
impersonate = "[email protected]"
billing_export_table = "my-app.billing.gcp_billing_export_v1_XXXXXX"

See Configuration for every key.

5. Check it worked

~/.local/bin/uvx platform-mcp doctor

Every environment should pass. Fix anything that fails before touching Claude Desktop — a server that cannot authenticate still starts, and then every tool call errors.

6. Edit the Claude Desktop config

Settings → Developer → Edit Config opens it, or edit it directly:

(Settings → Connectors lists hosted connectors; a local server like this one is not added there.)

The file usually already exists and holds your Desktop preferences. Add mcpServers as one more top-level key — do not replace the file, or you will lose those settings:

{
  "preferences": { "...": "your existing settings, left alone" },
  "mcpServers": {
    "platform-mcp": {
      "command": "/Users/YOU/.local/bin/uvx",
      "args": ["platform-mcp@latest"]
    }
  }
}

Replace /Users/YOU/.local/bin/uvx with what which uvx printed. On Windows the path looks like C:\\Users\\YOU\\.local\\bin\\uvx.exe — backslashes doubled in JSON. @latest makes each launch pick up new releases.

Check it still parses before restarting — a stray comma disables every server, silently. This uses only what ships with macOS:

osascript -l JavaScript -e 'function run(a){ObjC.import("Foundation");JSON.parse($.NSString.stringWithContentsOfFileEncodingError(a[0],4,null).js);return "valid"}' \
  ~/Library/Application\ Support/Claude/claude_desktop_config.json

7. Restart Claude Desktop

Quit it fully (Cmd-Q, not just closing the window) and reopen. Then ask "list the configured environments" — it should answer from list_environments.

If it does not appear, the logs are in ~/Library/Logs/Claude/ (mcp-server-platform-mcp.log). ENOENT there means the command path is wrong.

Why not a key file?

You can skip steps 2–3 by pointing GOOGLE_APPLICATION_CREDENTIALS (in the server's env block) at a downloaded service-account key. Think twice: a key is a long-lived credential sitting on a laptop, and the production read-only account can read every log, metric and billing row in that project. application-default login issues short-lived tokens tied to a person, and removing one grant revokes that person alone.

Observability

Every tool call appends one JSON line to ~/.local/state/platform-mcp/audit.jsonl:

{"ts":"2026-08-30T18:20:11+0800","tool":"query_logs","environment":"production",
 "project":"my-app","duration_ms":412,"count":50,"bytes":18422,"error":null}

Free-text arguments are recorded by name only — a Cloud Logging filter can carry user ids from the logs being searched, and the audit file must not become a second copy of that. Set PLATFORM_MCP_AUDIT_LOG to another path, or to off.

Diagnostic logs go to stderr (PLATFORM_MCP_LOG_LEVEL to adjust); in stdio transport stdout carries the protocol, so nothing else may be written there. In Claude Code, read them with claude --debug=mcp.

For a record that does not depend on this server at all, enable Data Access audit logs in GCP for the read-only service accounts. Token minting already appears in Admin Activity logs without any configuration.

Development

./.venv/bin/python -m pytest

The suite runs against an in-memory MCP client — no network and no GCP credentials — and covers environment resolution, the tool contract, annotations, error translation and the audit log. The only subprocess is bash -n over the packaged setup script, which is skipped where bash is absent.

The setup script lives at src/platform_mcp/scripts/ because it ships as package data; platform-mcp setup runs it from wherever the package is installed, so the quickstart needs no checkout.

Notes

• All tools cap result counts and truncate long payloads to stay token-friendly.
• GCP clients are built lazily and cached per environment, so switching between staging and production mid-conversation costs one client construction each.
• Cost recommenders are zonal/regional; list_cost_recommendations auto-discovers the locations where you have resources (via Asset Inventory) and fans out, skipping locations and recommenders that are empty or unavailable. It reports skipped_calls and fails loudly if it cannot discover any location, because "I could not look" and "there is nothing to save" must not look alike.
• get_cost_breakdown uses parameterized BigQuery queries with a whitelisted set of group-by columns, and filters to the selected environment's project. A billing export covers the whole billing account, so pass all_projects=true when you want account-wide totals.

License

MIT © 2026 Dimuthu Wickramanayake