01 The default
Most apps ask you to send your data somewhere.
A typical Marketplace app runs on its vendor's own infrastructure. Installing it means your issues, your customer requests and your page content are copied to a company you have not met, on hosting you did not choose, alongside sub-processors you would have to read a policy to name. None of that is unusual. It is simply how most of the ecosystem is built, and it is why every install turns into a review.
The data that cannot leak is the data that was never sent anywhere.
02 The constraint
So nothing is sent.
Everything here runs on Atlassian Forge. Atlassian hosts the code, the data stays inside the tenancy you already administer, and the app declares no egress — there is no address for your content to travel to, including ours. That is not a policy we could quietly change. It is a property of where the software runs.
Traffic circulates inside your tenancy. The region outside it is empty by construction, not by promise.
03 What it produced
Two apps built under that rule.
Both are on the Atlassian Marketplace, and the permissions each one asks for are on its install screen before you approve anything.
Mermaid Studio for Confluence
Flowcharts, sequence diagrams, ERDs, Gantt charts and mind maps written as plain text — with syntax highlighting, live preview, starter templates and one-click export.
Recurring Requests for Jira Service Management
Let portal customers repeat a request they have already raised — once, on a schedule, or from a saved template. Off until an admin enables it, with limits they control.
04 The same idea, elsewhere
Aion — batch work for coding agents.
Outside Atlassian the constraint looks different but the reasoning does not. Aion reads files and calls language models. It cannot run shell commands, and that is a decision rather than a gap.
Describe the job once. Run it over everything.
An agent asked to review forty files reads them one at a time, filling the conversation with material nobody will read and billing every file to an expensive model. Aion takes the mechanical part: one call per file, run at once, merged into a single answer. On a real forty-file review, 98.6% of what was produced never reached the conversation.
The wide pass runs on a cheap fast model and the merge on a capable one, results are cached against the exact prompt that produced them, and a spend ceiling stops a runaway job before it becomes a bill.
{
"prompt": "Review ${file.path} for SQL injection:\n${file.content}",
"files": ["src/**/*.py"],
"merge": "Summarize these findings, most severe first: ${results}"
}
One forty-file review, measured. 98.6% of what was produced never entered the context window.
05 Check it yourself
None of this needs to be taken on trust.
A promise not to misuse your data is worth less than not being able to reach it. Each of these is visible to you without our cooperation.
Hosting
Atlassian Forge. We operate no servers and hold no infrastructure of our own.
Egress
None declared. Nothing is sent outside Atlassian, to us or to anyone else.
Permissions
Listed on each app's page and on the install screen before you approve it.
Telemetry
None. No analytics, no tracking and no cookies — including on this site.