CloudPort Agent: Gemini-to-Qwen Cloud Migration Copilot

CloudPort Agent: production-ready Qwen Code Agent Skills that turn a 3-day Gemini→Qwen cloud migration into a 5-minute, human-approved autopilot workflow.

Comment

Inspiration

Alibaba Cloud holds ~4% of a $50B+ market. The barrier isn't price or model quality — it's migration cost. We experienced it firsthand: migrating our production app SoccerScope (a multilingual RAG agent for World Cup 2026 videos) from Gemini/Cloud Run to Qwen/Alibaba Cloud took 3 full days of manual work. We packaged every hard-won lesson into native Qwen Code custom Agent Skills, so the next team's migration takes 5 minutes.

What it does

CloudPort Agent is an Autopilot Agent (Track 4) that end-to-end automates a real-world business workflow: migrating live production applications from the Gemini/GCP stack to Qwen/Alibaba Cloud. It handles ambiguous, real-world codebases, invokes external tools, and enforces human-in-the-loop approval checkpoints before every mutation. It covers all four layers a real migration touches: LLM swap (Google ADK + LiteLLM → dashscope/qwen-plus), RAG re-indexing with Qwen text-embedding-v4 for multilingual vector search on MongoDB Atlas, MCP server integration (official MongoDB MCP) on serverless, and a structured-output compatibility layer for DashScope.

How we built it

Native Qwen Code custom Agent Skills, validated against a real production migration and deployed proof: SoccerScope now serves live traffic on Alibaba Cloud Function Compute 3.0. Architecture diagrams, evidence links, and the Apache-2.0 licensed source are in the repo.

Challenges we ran into

We initially considered using SAE (a service equivalent to Cloud Run) but ultimately decided to switch to Function Compute 3.0, which has a proven track record. Function Compute builds the cloud environment by uploading a ZIP file prepared in a local environment; however, we discovered that including MCP (Node.js) resulted in an excessively large package and caused npx to time out during cold starts, forcing us to temporarily abandon the use of MCP altogether. We subsequently resolved this issue by using the "Layer" feature in Function Compute to pre-install Node and rewriting the MCP startup process to directly invoke Node using node_modules bundled during the build. Solutions for these various pitfalls have now been codified as "skills" within our knowledge base, ensuring that other team members do not have to struggle with the same issues.

Accomplishments that we're proud of

Measurable impact: 3 days of expert manual migration compressed to 5 minutes with 8 human approval gates. A real production app — not a toy demo — running on Qwen + Alibaba Cloud today.

What we learned

Migration knowledge is perishable and expensive; Agent Skills make it reusable. Human-in-the-loop is what makes autopilot trustworthy in production.

What's next

More source stacks (OpenAI/AWS → Qwen), MCP-based verification tooling, and a public trap catalog (technical blog post).

Built With

  • alibaba-cloud-function-compute
  • dashscope
  • dashscope-(openai-compatible-api)
  • mcp
  • mongodb-atlas
  • mongodb-atlas-(vector-search)
  • openai-sdk
  • pydantic
  • python
  • qwen-code
  • qwen-plus
  • rag
  • text-embedding-v4
Share this project:

Updates