Home

Aura Tracker GCP

An open-source Model Context Protocol server that gives LLMs structured, safe access to your Google Cloud Platform infrastructure.

Safe by default — no configuration required:

• Read-only out of the box. 71 of 73 tools are pure reads. The two write operations require a valid preview plan and confirmation call before anything changes in GCP.
• Explicit cost controls. Chargeable BigQuery cost reasoning is off until COST_REASONING_ENABLED=true; each query is dry-run and bytes-bounded. Multiple environments can route to separate billing-export datasets without cross-environment fallback. Cloud Recommender can be disabled with RECOMMENDER_ENABLED=false.
• Two-step mutation gate. A state-drift-aware preview/confirmation protocol is enforced at the server level. The optional recommendation export uses the same gate. The server cannot prove that a distinct human initiated the second call; mandatory human review must be enforced by the MCP client or deployment policy.
• Authenticated remote access. SSE requires a Google ID token by default; authentication can be disabled only on a loopback endpoint.

→ Full details: Safety & Cost Safeguards

---

Why Aura Tracker GCP?

Managing distributed GCP infrastructure requires juggling the Cloud Console, gcloud CLI, and dozens of service-specific APIs. When something goes wrong at 2 AM, piecing together logs, metrics, IAM bindings, and traffic splits across ten browser tabs is slow and error-prone.

Aura Tracker GCP is a bridge. It exposes 73 default operations, one opt-in read-only cost-reasoning operation, and one opt-in BigQuery recommendation-export operation across the GCP service modules. Connect it to Claude Desktop (or any MCP-compatible host) and ask in plain language:

"Which Cloud Run services have an Aura Score below 50 and what's causing it?" "Show my production architecture." "Production is failing. Diagnose likely root causes and show me the evidence." "Why did costs increase this week? Show the top offenders, new resources, idle resources, and unexpected traffic." "Check yesterday's costs for preprod using its configured billing export." "Show me the Pub/Sub topics with unacknowledged message lag above 10,000." "I need a project security audit. Show Critical, High, Medium, Low, and the score." "Find configuration drifts between dev and prod, and say exactly which alias is missing each resource." "Scale the api-gateway node pool from 3 to 5 nodes — but show me a preview first."

The LLM calls the right tools and assembles the answer. For mutations, the server requires a valid preview plan and a second confirmation call before touching GCP; use a client approval UI or organizational policy when a distinct human must approve that second call.

---

The Problem

GCP's complexity grows non-linearly with team size. A medium project might have:

• 15+ Cloud Run services across 3 regions
• GKE clusters with 40+ workloads
• Dozens of Pub/Sub topics wired to Eventarc triggers and Workflows
• IAM bindings spread across hundreds of resources
• Observability gaps you don't know exist yet

Finding which service is responsible for a latency spike — tracing it from Cloud Trace through IAM to a Pub/Sub subscription — requires cross-product context a human has to assemble manually.

---

The Solution

┌──────────────────────────────────────────────────────┐
│  You (natural language)                              │
│         ↓                                            │
│  LLM (Claude, GPT-4o, …)                            │
│         ↓  MCP tool calls                            │
│  Aura Tracker GCP  ←──────────────→  GCP APIs       │
│         ↓                                            │
│  Structured answer + safety confirmation             │
└──────────────────────────────────────────────────────┘

The server runs locally (or behind an HTTPS endpoint in SSE mode). It authenticates to GCP using Application Default Credentials or a service account, while SSE callers authenticate separately with a Google ID token. Results are returned through your configured MCP host and LLM data path; enable the anonymization engine when sensitive values must be scrubbed before leaving the server.

---

Quick Start (3 Steps)

1. Install

Homebrew (macOS / Linux) — recommended

brew install asbrodova/tap/aura-tracker-gcp

Direct binary download (all platforms)

Download the archive for your platform from the latest release, extract, and place the binary on your PATH.

# macOS Apple Silicon example
curl -L https://github.com/asbrodova/aura-tracker-gcp/releases/latest/download/aura-tracker-gcp_darwin_arm64.tar.gz \
  | tar xz
sudo mv aura-tracker-gcp /usr/local/bin/

Building from source, Go toolchain install, and Docker usage are covered in Getting Started.

2. Authenticate with GCP

gcloud auth application-default login

3. Connect to Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "aura-tracker-gcp": {
      "command": "aura-tracker-gcp",
      "env": {
        "GCP_PROJECT_ID": "my-project"
      }
    }
  }
}

Restart Claude Desktop. Try: "List my Cloud Run services."

For multiple projects, configure named environments instead. Aliases are case-insensitive, and omitting one uses the default:

{
  "mcpServers": {
    "aura-tracker-gcp": {
      "command": "aura-tracker-gcp",
      "env": {
        "GCP_ENVIRONMENTS_JSON": "[{\"project_id\":\"my-company-123\",\"alias\":\"dev\",\"default\":true},{\"project_id\":\"my-company-345\",\"alias\":\"prod\"}]"
      }
    }
  }
}

You can then ask "Check logs on PROD" or just "List my pods". A configured project ID is also accepted in a request, but when that project has an alias, Aura Tracker returns only the alias. See Configure Your Environment for validation and privacy behavior.

---

What's Included

Category	Tools	Highlights
Compute	GKE (4), Cloud Run (6), Functions (2)	Aura Score, bottleneck detection
Data	Cloud SQL, Spanner, AlloyDB, Firestore, Memorystore	Read-only inventory
Messaging	Pub/Sub (3), Eventarc (2), Scheduler (1), Workflows (2), Tasks (1)	Health + lag inspection
Networking	Load Balancers, VPC, NEGs, API Gateway, PSC	7 tools
Observability	Monitoring (8), Logging (1), Coverage (1)	SLOs, traces, alerts, coverage gaps
Incident response	Incident Diagnosis (1)	Correlated deployments, metrics, IAM changes, logs, dependencies, platform health
FinOps	Cost Reasoning (1, opt-in), recommendation export (1, opt-in)	Historical comparison, ranked drivers, new/idle resources, traffic anomalies, BigQuery export
Architecture	Topology (1), Serverless Graph (1), Architecture Graph & Diagrams (2)	Live Mermaid, Graphviz, or SVG diagrams plus full graph export
Environment comparison	Drift Detection (1)	Whole-environment or component-scoped, alias-specific configuration differences
IAM	Permissions test, bindings, service accounts	3 tools
Security	Project security posture	Inherited IAM/deny, secrets, serverless and GKE exposure, effective firewall policy layers, KSA/GSA identity mappings, recommendations, score
Supply Chain	Artifact Registry, Cloud Build, Service Directory	4 tools
Advanced	GKE Workloads (5), GKE Mesh (1), Tagging (1)	Kubernetes-level visibility

73 tools by default; up to 75 when cost reasoning and recommendation export are configured. Full reference: Module Reference

---

Wiki Navigation

Page	What it covers
Safety & Cost Safeguards	Default safety posture, cost controls, mutation protocol, and the human-approval boundary
Getting Started	Prerequisites, install, auth, first query, module selection
Configure Your Environment	All env vars, --modules flag, SSE mode
Security and Safety	PII scrubbing, IAM least-privilege guide
Project Security Posture	One-call project audit, severity findings, score, recommendations, coverage, and architecture diagram
Managing GCP Costs	Paid APIs, billing alerts
Built-in Prompts and Workflows	audit-security-posture, optimize-bigquery-costs, incident-response-helper
Incident Diagnosis	gcp_incident_diagnose, evidence scoring, scope safety, dependency checks
Drift Detection	gcp_compare_environments, supported components, alias semantics, normalization, coverage, and examples
Cost Reasoning	gcp_cost_explain, billing-export setup, drivers, safeguards, and limitations
Recommender BigQuery Export	Opt-in recommendation snapshot, IAM, dry-run, confirmation, and retry behavior
Automatic Architecture Diagrams	Production scoping, Mermaid, Graphviz, SVG, parameters, and examples
MCP Resources Reference	10 built-in resources (BigQuery, Cloud Run, Storage, IAM)
Module Reference	Index of 30 default modules plus optional recommendation export
Architecture and Contributing	Hexagonal architecture, adding tools