An AI product shipped and run from Claude Code
The team connects Faable's MCP server once, then deploys, reads logs and adds domains by asking. They never leave the terminal, and the agent never holds the keys to everything.
Illustrative scenario based on patterns we see across teams building on Faable.
The team
A team of three building a support assistant for online shops on top of the Claude API. They write their code with Claude Code and wanted to ship and run it from the same place.
The challenge
Opening a hosting dashboard for every deploy, log or domain broke the flow of a team that already works through an agent. Giving that agent access to infrastructure raised the obvious questions: can it delete things, and can it reach other projects? LLM apps have their own traps too: slow first tokens that proxies cut off, work that doesn't fit in a request, and memory lost when it lives on disk.
How they built it
- 1
Faable's MCP server in Claude Code
One command adds it, and a consent screen asks which single project the connection may act on. There is no API key to copy.
- 2
Deploy by asking
"Deploy this repository to Faable and give me the URL." Claude Code creates the app and waits until it's live. From then on, every push deploys.
- 3
Operate by asking
Claude Code reads the build log of a failed deploy and says whether the cause is the code or Faable. It also reads the runtime logs and attaches a domain, returning the CNAME to create.
- 4
An agent with limits
The hosted server acts on one project, has no tool that deletes an app, and has a read-only mode. Logs reach the agent marked as data, not instructions, and real secrets are set from the team's own terminal.
- 5
Built for LLM workloads
The FastAPI app streams Claude's answers over Server-Sent Events (the limit is 60 seconds to the first byte, not to the last). A background worker drains the job queue, and a scheduled task builds the nightly reports. State lives in a database, because the filesystem is ephemeral.
The result
Deploys, logs and domains without leaving the terminal. The agent's reach is limited to one project with no destructive tools, and the app is hosted in the EU.

