Vibe guide

Describe the app you want in your own words. Ocean Vibe asks what it needs to know until the brief is complete, then a crew builds the app on Plunge's own parts: screens on App UI, data on Ocean Base, automations on the platform's workflows and schedules. Every round is checked, reviewed and tried before it reaches you, you accept it at the door, and every landed round can be undone. You watch it all happen in one live workspace.

Start an app

1. Open vibe.plungeai.com (opens in a new tab) signed in to Plunge (you are sent to sign in otherwise).
2. Under New app, give it a title and say what it should do. A sentence is enough ("a service desk for our IT team"); a pasted document works too. A message holding a key (an ozk_ key, a private key) is refused and nothing of it is kept: remove the key and send again; keys belong in Connections.
3. Ocean Vibe asks about one thing at a time: who uses it and for what, what it keeps, its screens, what should happen on its own, who may open it, how you will know it works. Each question proposes an answer; say yes or say what is different.
4. When nothing is missing, the workspace shows the brief card: the app in a few lines. Read it; Change something if it is not right, Looks right, build it when it is. Then the workspace offers Build.

open on the brief lists what you asked for that the platform cannot do yet (for example "text the on-call agent by SMS"): it is kept in the brief and left out of the build, so you know before anything is built.

Rounds

Build starts a round. It runs on its own through eight stations, listed in the Round rail on the right with a dot and a word each:

1. plan: a desk reads the brief and writes what will change.
2. build: a desk writes the app as YAML: its data, its screens, its example rows and its automations.
3. check: Ocean Vibe's own checks, no model. A finding sends the round back to build with the finding as the note.
4. try: your app's database gets a branch for this round (<name>@round-<n>), the schema is applied on it, the example rows are added, a draft of the app is published and every screen is opened.
5. verify: three desks, each on a different model from the builder's, at once: review (does it do what the brief says), security, and acceptance, which uses the draft the way an agent would and tries each of the brief's acceptance lines.
6. door: you decide (below).
7. land: the branch becomes the database, the app is published, the automations are saved.
8. watch: for a few minutes every screen, schedule and webhook is checked; a problem is said in the chat.

A round may go back to build once; when it would go back a second time, or when something it cannot fix stops it, it parks and the chat says why. Retry round starts a parked round again; Cancel round stops one that runs. A desk may ask you something in the chat; your answer starts its station again. One round runs at a time: what you say while one runs goes to the next round. A round may spend 3 USD, a project 25 USD and 10 rounds a day.

The door

When a round is ready the door card appears above the stations. It says first what the round changes: it drops data or changes who may open the app, it sends something outside, it adds, or it changes words and order. Beside it you have the preview of the draft, each screen as agents read it, the schema plan, the changed files, the three verdicts, the acceptance lines and the cost.

• Accept lands the round.
• Request changes asks for a note and sends the round back to build with it.
• A round that drops data lists what it drops and shows Allow this round to drop data; Accept needs it ticked. When the drop is found while the draft is being made, the card comes early and the round waits for the tick.

If the database changes while you look (someone adds a row), the land does not force it: the round takes a fresh branch, tries again and comes back to the door.

Undo

Every land keeps where the database was before it and the app version it replaced. Undo last land, within 24 hours, puts the app back to that version, rewinds the database and wires the earlier automations again, as one action; the chat names anything that could not come back (a webhook trigger the land removed comes back as a new one). It is refused while a round runs, except during the landed round's own watch. Undo last land sits in the Round rail while an undo is possible, and asks before it rewinds the live database.

The tabs

• Preview: the app itself, the round's draft while it has one, else the published app, with a screen switcher. Before the first draft it shows the brief card.
• Screens: each screen as agents read it (its Markdown), the action lines kept.
• Data: the schema applied, this round's plan (a step that drops data marked as such), what the round's branch changed, and the rows of a table you pick.
• Files: the project's documents, each one a link, and this round's changes line by line.
• Automations: each automation with its trigger, its saved workflow, its last run and its next.
• Activity: what happened, newest first: stations, desks, checks, verdicts, the door and the land.

Running desks on your Mac

You can run a round's desks with your own oceancode and your own models instead of the platform's.

1. Open vibe.plungeai.com/token (opens in a new tab) and press Regenerate; the desk token is shown once.
2. Run vibe token and paste it; it is kept in your Mac's keychain, never in a file.
3. Press Build on my Mac beside Build (Build itself starts platform desks). The round's plan, build and verify stations then wait for your desks.
4. At each station the rail shows the command to run, vibe desk --project <vp_> --round <vr_> --role <role>, with a Copy button; paste it into your terminal (--model author/model and --budget-usd are optional).

A desk ends with VIBE_DESK_OK role=… cost_usd=…, and the rail shows its result like any other desk's. A station whose desk is not started within 30 minutes parks the round, and a desk quiet for 30 minutes is ended, so a desk that died never holds the round.