Mini-apps (Jenny Apps)

A Jenny App is a small app with its own screen that Jenny builds for you on request — a shopping list, a counter, a custom dashboard for a device on your network.

Apps vs. skills

Jenny has two ways of gaining a new capability, and they answer different questions:

If it needs a screen, it’s an app. A Jenny App has its own UI you open from the Apps grid: a list you check off, a form you fill in, a chart you look at.
If it only lives in chat, it’s a skill. A skill just teaches Jenny a procedure — there’s nothing to open. See Skills.

An app can still talk to Jenny (see Directionality below), and it can integrate with an external server, but the defining trait is: does the user need to look at a screen for this, or does it just happen in conversation?

Creating an app

Apps are created only through chat — there is no dedicated app editor in the WebUI. Two ways to start:

Just ask, in your own words: “make me an app for tracking my plants.”
Tap the + tile in the Jenny Apps section of the Apps tab. It opens the chat with this guided prompt already typed:

I want to create a new Jenny App. Use the “app-creator” skill and guide me step by step.

Jenny then walks you through the design conversation (what the app tracks, what actions it needs) using the built-in app-creator skill, confirms with you, and writes the files. Editing an existing app works the same way: long-press the app tile and choose Edit — this sends the chat prompt I want to edit the Jenny App "{name}" (slug: {slug}). Can you help me? rather than opening any in-app editor.

Where an app lives

Every app is a folder under workspace/apps/<slug>/:

Path	Contents
`app.json`	The manifest: name, icon, an optional external server, and the list of typed actions.
`AGENT.md`	Context for Jenny about the app — what it’s for, preferences, thresholds. Not part of the technical contract, just briefing notes.
`app/index.html`	The UI. HTML and JavaScript only — there is no per-app Python. Only this `app/` folder is served over the web; the manifest, `AGENT.md`, and `data/` are never reachable by URL.
`data/`	JSONL collections that hold the app’s data, shared between the app and Jenny.

Because it’s a normal folder in the workspace, you can also open and edit these files by hand from the Workspace tab if you want — Jenny doesn’t have to be the one writing them.

Every action is also a tool

Each action declared in app.json becomes a native tool that the LLM can call directly, named <slug>_<action> — for a “plants” app with a water action, that’s plants_water. These tools are re-registered on the fly whenever app.json changes, with no restart needed.

This has a consequence worth calling out explicitly: Jenny can read and modify an app’s data even while the app is closed. A closed app is just unrendered HTML; nothing stops Jenny from appending a record to its data/ collection because you asked her to in chat. The reverse is not true — see below.

Jenny and the app talk in one direction

Jenny → app: always possible, whether or not the app is open on screen.
App → Jenny: never on its own. The app can only hand off to chat through an explicit user action, jenny.discuss(), which switches you to the chat view with the app’s context pre-filled. Jenny’s replies always land in chat — they are never rendered inside the app itself. The app stays a deterministic screen; the thinking happens in one place only.

What an app can and cannot do

Apps run inside a sandboxed iframe with allow-scripts only. That means:

No native <form> submission, no alert()/confirm()/prompt(), and no direct fetch() calls out of the app’s own JavaScript. All data access goes through the app SDK’s jenny.action() call.
The app cannot navigate the rest of the WebUI or touch the chat DOM.

Storage actions (append/set/update/delete/query on a data/ collection):

Limit	Value	What happens beyond it
Collection size	5 MB per collection	Writes fail with HTTP 413.
Query result size	200 records by default	Records are read oldest-first and the list is cut at the limit, so once a collection passes it, the newest records are the ones silently left out — not the oldest — unless the app asks for a higher limit.
Malformed line in a collection	—	Skipped silently; only a warning is written to the log.
Updating/deleting a record that doesn’t exist	—	Fails with HTTP 404.

Actions that call an external server (a http-kind action, e.g. talking to a LAN device):

Limit	Value
Request timeout	20 seconds
Response size	512 KB
Reachability	LAN addresses are allowed; loopback and cloud-metadata addresses are blocked
Redirects	Never followed
Parameters (GET, ~6 KB budget)	Requests with a larger encoded parameter payload fail with HTTP 413

Authenticated external servers are not supported. An app’s manifest can describe a server.auth credential reference, but there is no credential store behind it. If an app declares server.auth, every call to that server is refused fail-closed with HTTP 501 (“no credential store is configured”) rather than being sent without credentials. Don’t build an app around the assumption that it can log in to a service on your behalf — it can only talk to servers that don’t require authentication (e.g. a plain LAN device).

When an app is broken

If a manifest fails to load — malformed JSON, an invalid action definition — the gateway never crashes. The app simply shows up in the Jenny Apps grid tagged broken, with the readable error underneath its name. Tapping it prompts: The app "{name}" is broken: {error}. Ask Jenny to fix it? — confirming sends the error straight to chat so Jenny can look at the files and repair them. Since the app generator is itself an LLM, occasionally getting a manifest wrong is expected, and this is the recovery path.

Deleting an app

Long-press an app tile and choose Delete. The confirmation reads Delete app "{name}"? It will be removed permanently. — and it means it: this deletes the whole workspace/apps/<slug>/ folder, including its data/. There is no trash or undo from the UI. Your only safety net is the automatic workspace snapshot history, which is not something you can browse per-app — restoring one means restoring the entire workspace to an earlier point in time.