Overview

How OpenApps Network works.

OpenApps is a family of open-source apps that share one account, one credit balance and one identity, and that AI agents can use as well as people. This page shows how the pieces fit, what runs today, and what we are building next.

Last updated: 5 October 2026

Four pieces, one network.

Each app does one job well: edit a PDF, read text from a scan, subtitle a video, upscale a photo. Under them sit three shared services, so that a person signs in once, pays once, and can hand the same work to an agent.

What happens when an agent does a job.

Say you ask Claude to split a folder of scanned invoices into one PDF per invoice and pull out the totals. With OpenApps MCP connected, this is the path the request takes.

1. Find. The agent searches the catalogue in plain words, in any of our eight languages. It gets back a few compact cards, each with a price range, rather than every tool we have.
2. Plan. It hands over the job and the inputs. OpenApps returns a plan with a quote, what the plan will read and write, and how sure we can be of the result.
3. Approve. OpenWallet checks the quote against the limits you set. Within them, the run goes ahead. Above them, you get a prompt from OpenWallet on your own device, never in the chat.
4. Run and watch. The work runs in the app that owns it. The agent polls for progress, or a screen can watch the run live.
5. Result and receipt. The agent gets download links for the results and a signed receipt it can check without asking us. If the run fails, the system refunds it; the agent does not have to.

Today, step 3 already works for payments through OpenWallet's own connector. The other steps arrive with OpenApps MCP; the agents guide marks which is which.

The rules we build to.

• Your files stay yours. Apps work on your device wherever they can. When work has to run on our servers, files move as upload handles and short-lived links, never pasted into the conversation, where they would end up in logs and in the next model call.
• The model never approves its own spending. Every operation declares what it does: read, write, send or spend. OpenWallet decides on that declaration, outside the model, and anything over your limit waits for you.
• You pay for what worked. Credits are charged by the server that did the work, once it worked. Every charge carries an idempotency key, so a retry cannot charge you twice.
• Every paid run ends in a signed receipt. Receipts are signed with Ed25519, and the public keys are published, so anyone can check one offline.
• One door, not fifty tools. An agent connects to one address with six tools and searches for the rest, so it isn't paying for our whole catalogue on every turn.
• Built in the open. The apps are open source, and so are the payment library and the desktop wallet agent.

What runs today, and what is next.

As of 5 October 2026. We update this table when something ships.

Where to go next.