Having one’s own vibe coding agents seems to be a must-have for distribution of PaaS/DBaaS: see Replit, Vercel, Neon, Convex, and many others. Fly.io is no exception, with phoenix.new.
However, for me as a practicing dev, a much more attractive platform capability would be a DevOps agent that knows the platform’s (that is, Fly.io’s) capabilities, APIs, and deploy config format (fly.toml) by heart, and can (try) deploy a specific open-source Github project for me.
Not “What will you build today?”, but “Give me a Github link, I will figure out how to deploy the system on Fly.io and will discuss tradeoffs with you”. Tradeoffs such as: this specific projects seems to require a vector DB for functioning, but it only mentions Pinecone and Qdrant in the description. You can either register with one of these providers and give the agent the access creds, or the vibe coding agent may try to change the project’s code to support LanceDB or Turbopuffer and get away with object storage that is included in Fly.io’s bill.
I love OSS and I love Vibe coding (full disclosure: the term “Vibe coding” still doesn’t thrill me, but the practice is something I now use heavily); but that scope is too broad to make meaningful progress on.
Let’s talk a bit about fly launch and fly mcp server and explore what the next steps should be.
Today, in an empty directory, you can run a command like:
… and (without any AI agent) it will try to do what is necessary to make this work. For many projects, this just works. Others will need some adjustments or what you call “tradeoffs”.
Now lets talk about fly mcp server. This will enable the AI agent of your choice (Claude, Code, Cursor, Zed, whatever) to interact with fly.io. It doesn’t yet support fly launch for the simple reason that I haven’t gotten around to it yet. The description for the command could contain preconditions or post conditions - things to check before launching, and things to check after launching, in particular after a failed launch.
As I said, I haven’t gotten started on this work, but that’s mainly because I don’t know where to start. OSS is too big as a first step. But if I can get a stream of reports along the lines of “I tried fly launch with this project and got stuck here”, I can make fixes either to fly launch or fly mcp server and iterate.