AI agent
From an instruction to a finished task.
Describe what you want in plain words. The agent reads your open tabs, uses the apps you have installed and works in your Library. Steps that change or send something need your approval, and you decide how often it has to ask.
Agent panel · Task
“Read the rates notice I have open and note the due date and the amount in my Library.”
- Read the notice in the current tabDone
- Found the due date and the amount payableDone
- Saved “Council rates — due 28 Oct” to your LibraryYou approved
Approval needed
Save a note to your Library
Saving is low risk, but this is the first time this task has written to your Library, so it asked.
Result
A note with the amount and due date is in your Library. Nothing was paid; payments are always yours to do.
Approvals
You approve the important steps.
How often the agent asks depends on what a step does. The level is worked out by the browser; a task or an app can never lower it.
| What a step does | Examples | What happens |
|---|---|---|
| Reads and savesReading a page or your Library, saving a note, using a model on your device. | Reading a page or your Library, saving a note, using a model on your device. | Runs without asking. |
| Changes you can undoChanging something on a website that can be changed back. Sending what it read to your own cloud model. | Changing something on a website that can be changed back. Sending what it read to your own cloud model. | Asks you. You can allow it on that site from then on. |
| Things that cannot be taken backSending, posting, submitting or deleting on a website. Sending data to a declared destination. | Sending, posting, submitting or deleting on a website. Sending data to a declared destination. | Asks every time, until you have approved the same step three times on that site. |
| Never done by the agentPayments, passwords and account security. | Payments, passwords and account security. | Handed back to you, always. |
A step is treated as more sensitive when it works from text the agent read on an untrusted page, when the task comes from an unreviewed author, or when it is the first time it runs.
Tools
Work across browsing, apps and your Library.
Your open tabs
The agent can read pages you have open and act on websites within the approval rules above. It does not browse the web on its own account.
Installed apps
An app can offer tools to the agent, so “add this to my budget” can hand the figures to Budget and “file this receipt” to Receipts.
Your Library and chosen files
It can read and save in your Library, and open files only from a folder you have picked for it. Nothing else on your disk is visible to it.
What it can do
More than a chat box.
The agent lives in a panel beside the page. It can answer a question about what you are reading, or go and do something.
- Ask about the page
- Summarise a long article, pull the figures out of a table, or explain a clause, using the page you have open as its source.
- Fill in forms
- It types into fields, picks from menus, ticks boxes and sets dates, then checks that the values took.
- Follow links and compare
- Open the pages a task needs, read them, and bring back what you asked for.
- Use your files, by invitation
- Point it at a folder and it can read what is inside. It sees nothing else on your disk.
- Its own window
- Ask for a separate agent window and watch it work without giving up the tab you are in.
- Connect other tools
- Add connectors and MCP tools you trust, and the agent can call them under the same approval rules.
- Skills
- Import a skill, a set of written instructions for a kind of task. A skill guides the agent; it never grants extra permissions.
- Saved workflows
- Keep a task that worked and run it again with new inputs.
- Build an app
- Describe a tool and the agent writes a small app for your browser.
Models
Choose how your AI runs.
OwnWeb does not come with one fixed model. You decide where the thinking happens.
On your device
Run a model locally, through Ollama or another compatible service. What the agent reads stays on your computer, and it costs nothing to use.
Needs a capable computer and a one-time model download. Speed and quality depend on the model you choose.
A provider you connect
Add your own key for a cloud provider, or use Claude Code or Codex if you already have them installed. Your key is kept in the operating system’s key store.
Each provider needs your consent before anything is sent, and is used only after a short test passes for its address and model.
You can mix them: a small local model for routine steps and a stronger one for planning. Which models are offered, and how well a task goes, depends on what you connect.
Recovery
When a task needs your help.
Tasks do not always go to plan. The agent is designed to stop and tell you, not to guess.
- Stop at any time
- Every running task has a Stop control. Stopping leaves the work done so far where you can see it.
- It asks instead of guessing
- When a page does not match what the task expected, the agent hands back with what it found rather than improvising.
- Approvals time out safely
- An approval the agent is waiting on expires if you do not answer, and the step is not taken.
No agent is perfectly autonomous, and no agent works on every website. OwnWeb’s is built so the unexpected ends in a question to you, not an action you did not ask for. Passwords, CAPTCHAs and payments are always yours to do.
Next