Devin can now build private internal apps your team opens from a link

Cognition has launched Devin Apps, private web applications that its Devin agent builds and hosts on its own cloud computer. A team describes the tool it needs, and colleagues open it from a link. Every viewer must sign in to Devin, and Apps are never public.

How a prompt becomes a hosted tool

Per Cognition’s post, a user opens Apps in the sidebar, describes what the App should do, and clicks Build with Devin. Prompt templates are available, including a metrics dashboard, a project status page and a review queue. Devin starts a session, asks any questions in its first message, then builds the App and puts it online. The App appears on the Apps page at once with its URL and shows a Building status until the first version is live.

An App can also be requested from any session. If Devin has already shown a browser preview, the user can ask it to publish that preview as an App.

Because Apps run on Devin’s machine, Cognition says they can reach a team’s existing MCPs, integrations and codebase. They can use an internal tech stack and components, or begin from a built-in template. Any stack that serves HTTP on a port works, and an App can use a database such as SQLite or Postgres on the same machine. It can also call the Devin API.

Sharing, versions and the sleeping machine

Each App keeps a stable URL. The Share button lets an owner restrict access to named people, open it to the organization, or open it to the whole enterprise. Users can pin favorites, and organization admins can pin an App for everyone.

To change an App, the user clicks Edit in session and tells Devin what to alter. Devin builds a new version at the same URL, and the App’s page lists versions under Deployments. If a new version fails to start, the previous one keeps serving. An automation can also refresh data on a schedule or redeploy when a branch changes.

Apps rest on memory snapshots, so the machine hibernates when nobody is using the App and wakes on the next visit. Cognition says the first load after a sleep can take a few seconds, and that billing follows how long the machine is awake. Cognition also states the limit plainly: an App runs on one machine, which suits internal tools and dashboards but not high-traffic production services.

The company lists three uses: internal tools such as admin panels, request forms and approval queues that replace a shared spreadsheet; dashboards that pull from a team’s own data on a schedule; and beta or staging copies of a product, with frontend, backend, database and workers, that teammates open from a link.

Billing by awake time shifts the cost question

The post gives no price for Devin Apps, no per-hour rate for awake time, and no plan restrictions. That is the main gap for anyone budgeting. What it does say is that cost follows machine uptime, not a seat or a hosting tier. For a tool used a few times a day, the machine sleeps most of the time, which is the scenario the design targets. For a dashboard that refreshes on a schedule or is opened constantly, awake time could add up, and the source does not quantify it.

Devin’s own work is already metered per task. Our coverage of Gemini 4 Hydrogen on Devin reported 38.2% on Devin’s coding test at $2.40 a task, and our report on Gemini 4 Carbon put that model at 40.0%, up from 26.3%. Those figures describe the models Devin runs, not Apps, but they show the scale of per-task spending that building an App would sit alongside, with hosting time billed on top.

The comparison Cognition draws is a shared spreadsheet or a request that waits in an engineering backlog. An App built in a session replaces that wait with a single task, though the post offers no data on how long a build takes or how often the first version works. Those claims are the company’s own and have not been independently verified.

What the announcement leaves open

Access is tied to Devin sign-in, so every reader of an internal dashboard needs a Devin account within the chosen scope. That is a security advantage, since nothing can be exposed publicly, and a cost question if many occasional viewers need seats. The post does not say how viewer accounts are licensed, which regions or plans have access, or how much data an App may hold on its single machine. It points readers to the Devin Apps page and the docs for more.