Recipes: publish a project, deploy it with one click

A recipe is a snapshot of a project you own, published so that anyone can deploy their own copy from a single link. The copy is a new project in their account, paid for by them.

What a recipe is

A recipe is the shape of a project at one moment: its services with the images they run, its scheduled jobs, its resources, and the names of the external credentials it needs. It is stored as immutable versions. Nobody who deploys it ever touches your project, and your project never learns who did.

Publishing

gg recipe snap PROJECT --dry-run
gg recipe snap PROJECT

Only the project’s owner can snap it; an editor gets owner_only. Run it with --dry-run first: it prints the manifest, the version it would publish and every refusal, and changes nothing. The images are copied into a registry space the recipe owns, so a later change to your project cannot alter a published version.

What a snapshot carries

  • Services, with their image, size, port, volume and plain environment variables.
  • Scheduled jobs.
  • Resources: postgres, qdrant, valkey, iggy, rabbitmq, and the edges between them.
  • External resources, as inputs by name only: the variable names are recorded, and the values never are.

What it drops, in v1:

  • Custom domains. They belong to whoever owns the DNS.
  • One-shot jobs. Nothing runs at install time.
  • S3 buckets. Snap refuses a project that has one.

Env never carries secrets

Plain environment variables travel with the recipe, so anything secret-shaped in them stops the snap with env_looks_secret. Secrets belong in an external resource, which the recipe keeps only as a list of names. This is a rule, not a suggestion: a recipe is readable by everyone who has its link, and a value that went in cannot be taken back out of a version that someone has already installed.

Inputs

Whoever deploys the recipe supplies the values for its inputs. On the CLI they come from env files, never from argv, where they would land in shell history:

gg recipe install ID my-copy --input openai=./openai.env

On the deploy page they are masked fields. Either way they are stored as the deployer’s own secrets.

Versions

The first snap is 1.0.0. After that snap works out the lowest version your change allows, and refuses a lower one:

What changedAt least
Something removed or renamed, or an immutable field changedmajor
Something added, or anything that raises the deployer’s billminor
Otherwise, such as a new image digestpatch

The link and the badge

Every recipe has an eight-character id, and its deploy page is https://my.gagarin.cloud/deploy/<id>. To put a button in a README, paste:

[![Deploy to Gagarin Cloud](https://my.gagarin.cloud/deploy/<id>/badge.svg)](https://my.gagarin.cloud/deploy/<id>)

An install always creates a new project. It never writes into one that exists, so deploying a recipe cannot overwrite anything.

Updating an install

gg recipe diff PROJECT
gg recipe update PROJECT

diff shows what moving to the newest version would change. update applies it under the rules every deploy follows: a new image is a deploy, additions are free, a new input is asked for, and a removal needs the owner’s approval. A field the deployer changed since installing is kept, and if the recipe changed the same field the update refuses unless you pass --force.

Closed source is fine

A recipe ships images, not source, and source_url is only a link. Recipes are unlisted by default: they work for anyone with the id and appear nowhere else until a public directory exists.