
GCP Platform Engineer
io.github.deBilla由 GitHub 帳號 deBilla 發布v0.4.0更新於 Oct 24, 2026
Read-only GCP tools for debugging incidents and finding cost savings
概覽
唯讀的 GCP 伺服器,讓助理檢視日誌、指標、資源與帳單,用來排查事故並找出成本節省空間。
- 可以做什麼
- 列出並切換已設定的 GCP 環境,結果會回報所屬環境與專案
- 查詢日誌、取得近期錯誤並彙整錯誤群組
- 查詢指標、列出警示政策與可用性檢查
- 檢視成本明細、帳單資訊以及成本與其他建議
- 搜尋雲端資源,並盤點 Compute、Cloud Run、GKE 與 Cloud SQL 執行個體
- 適用情境
- 適合讓助理排查 GCP 事故、盤點資源,或在預備環境與正式環境之間尋找降低成本的机会,同時不讓它具有變更基礎架構的能力。
- 執行需求
- 本機執行的 Python 3.11+ 程序,通常透過 uvx 或 pipx 啟動。需要一個 Google Cloud 專案與 Application Default Credentials(個人登入或服務帳號),一次性設定需要 gcloud CLI,並在各專案具備檢視者角色及帳單檢視權限。環境可透過 TOML 設定檔或 PLATFORM_MCP_ENVIRONMENTS 變數設定;成本明細另需 BigQuery 帳單匯出資料表。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 GCP Platform Engineer,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
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
gcloudCLI for the one-time setup
Install
From a checkout, for development:
Check your setup
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:
2. Grant read-only access to the identity the server runs as.
For local development with your own login (Application Default Credentials):
The identity needs these viewer roles on the project, plus roles/billing.viewer
on the billing account:
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:
Or by hand:
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:
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.serviceAccountTokenCreatoron eachplatform-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:
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:
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 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:
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:
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):
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:
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:
3. Authenticate
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
See Configuration for every key.
5. Check it worked
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:
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:
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:
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
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_recommendationsauto-discovers the locations where you have resources (via Asset Inventory) and fans out, skipping locations and recommenders that are empty or unavailable. It reportsskipped_callsand 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_breakdownuses 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 passall_projects=truewhen you want account-wide totals.
License
MIT © 2026 Dimuthu Wickramanayake
來源:README.md,提交 969e4ef
工具
0版本歷史
1- v0.4.0最新Sep 16, 2026
評分
已滿足適用檢查項的 50%
該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。
- 近 30 天內可連通25 分
請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。
- 已宣告認證方式10 分
請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。
- 已列出工具15 分
請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。
- 描述必需15 分
- 圖示10 分
- 文件15 分
- 授權條款10 分