migration-to-gcp

AWS-to-GCP Migration Skills Harness

AI-driven migration assistant that turns a snapshot of your AWS workload — Terraform, CloudFormation, application source, or billing data — into a reviewed, costed, and ready-to-apply Google Cloud migration plan.

Built as a plugin for Claude Code — bundling the migration skill, two official knowledge MCP servers (Google Developer Knowledge, AWS Knowledge), and a live GCP pricing CLI (gcp-pricing).

---

What this skill does

It drives Claude Code through a deterministic, multi-phase migration workflow that mirrors how a senior cloud architect would assess and plan a real AWS-to-GCP move:

Discover → Clarify → Design → Estimate → Generate

Each phase has a hard gate: the agent cannot skip ahead. Clarify must finish before Design, Design before Estimate, and so on. This is enforced by .migration/<run-id>/.phase-status.json, which the skill writes after every phase. The result is a reproducible, resumable run instead of a one-shot prompt that hallucinates a plan.

What the agent actually does behind those phases:

• Reads your sources directly — Terraform .tf files, CloudFormation / SAM / CDK templates, app source for AWS SDK call sites, and AWS Cost & Usage Reports — using the model's built-in Read / Grep / Glob tools. No parser scripts are emitted.
• Runs live research before every non-deterministic decision — pricing, service capabilities, lifecycle stage, latest GCP product fit — through the bundled gcp-pricing CLI, the google-developer-knowledge and awsknowledge MCP servers, and WebFetch on cloud.google.com. Training-data recall is explicitly not allowed for cost figures or service recommendations.
• Asks before guessing — discovery ambiguities (e.g. a Kinesis Firehose with no destination), migration intent (cutover / PoC clone / hybrid), cross-cloud connections, and any 2+ viable GCP target choice are surfaced to the user as multiple-choice questions, not silently resolved.
• Defers what it shouldn't auto-map — Redshift, for example, is always marked Deferred — specialist engagement because a sound BigQuery / Dataflow / Dataproc target requires a specialist conversation with your Google Cloud account team or a data analytics partner.

---

When to use it

Use it when you have at least one of these in hand and want a concrete migration plan to start from:

• Terraform (.tf, .tfvars, .tfstate) describing AWS resources.
• CloudFormation / SAM / CDK-synthesized templates (.json, .yaml, .yml, .template).
• Application source with aws-sdk-*, boto3, or @aws-sdk/* imports — gets you per-call-site SDK rewrite patches.
• AWS billing exports — Cost & Usage Report, Cost Explorer CSV / JSON, or any *billing*.csv / *cost*.json.

Typical scenarios:

• You inherited an AWS workload and need a costed re-platform plan for a leadership review.
• You're evaluating a PoC on GCP and need a parallel deployment generated from live AWS IaC.
• You're migrating LLM workloads from Bedrock / OpenAI to Vertex AI Gemini and want per-call-site rewrite patches.
• You want a defensible monthly-cost comparison (AWS → GCP) based on live Cloud Billing Catalog API prices, not stale rate cards.

When NOT to use it

The skill explicitly refuses or yields to specialists in these cases:

• Azure or on-prem → GCP — out of scope (AWS source only).
• GCP → AWS reverse migration — out of scope.
• General GCP architecture advice without migration intent — use the underlying google/skills directly.
• AWS-to-AWS refactoring — wrong tool.
• Multi-cloud deployments that do not involve migrating off AWS — wrong tool.
• Redshift → analytics warehouse — marked Deferred — specialist engagement; you're directed to your Google Cloud account team or a data analytics migration partner.

---

How it works

The phases