abox.tools

Language: English

Can an AI agent use these tools?

Agents can use the browser interface. To assess where data goes, check where the browser runs, what the agent itself sends and which tool features use the network.

View as Markdown

The short answer

An agent with browser control can use these ordinary web pages: labelled controls, file pickers and downloads, without a site account. Its environment must support the required file access, browser APIs and permissions. Camera permission or a device share sheet may need your participation.

For file processing, the first question is where the controlled browser runs. Also check what the agent's own service receives, and whether the requested feature intentionally fetches or shares data. A local browser alone does not settle all three questions.

Give the agent a job it can check

Preparing application attachments. Give the agent the required dimensions, accepted format and maximum bytes for each file, then ask it to keep the originals and return separate finished copies. A useful brief is: “Make these three photos JPEGs under 200 KB each; show me the dimensions, actual sizes and readable previews before I submit them.” A successful download alone does not establish that the form will accept it.

Preparing an expense packet. An agent can operate the receipt workflow, but the merchant, date, currency, total and crop still need review against each picture. Ask for the report and attachments to be prepared for your check. Choosing a recipient or sending them is a separate action.

The same approach works for a repeated monthly job: reuse the outcome and checks in the brief, with the new files. Confirm the agent's browser location and data handling each time the environment changes.

Two kinds of agent, one distinction

Agents that use tools come in two shapes, and the difference between them matters more than anything else on this page.

A locally controlled browser reads and processes the file on that device. Using a local processing tool here adds no file upload for processing. The agent may still send instructions, screenshots or file context to its vendor; that is governed by the agent's own implementation and settings.

A cloud browser runs on the agent vendor's computer. If you attach a file for the agent to use there, it has already left your device. These tools process it inside that remote browser and do not undo the earlier transfer. Intentional sharing or app handoffs are additional actions in either environment.

Specify a local browser when keeping processing on your device matters, check the agent's own data handling, and authorize any deliberate transfer separately. Online rate lookup and email handoff, image URL import, sharing and camera viewing have documented network boundaries.

How to hand a job to an agent

Agents do best with the same brief a colleague would want: the tool, the file, and what done looks like. Some patterns that work:

What not to delegate

Operating controls is only part of a task. Review decisions that change what another person can see or receive. Two examples:

Deciding what must not be seen. The redaction tools delete what you cover — but choosing what to cover is the job, and an agent that misses one line has produced a file that looks finished and is not. Let an agent operate a redactor if you like; look at the result yourself before it goes anywhere, the same rule those tools' own guides give a human operator.

Opening what was read. This site's QR reader never follows a decoded address automatically, because reading and following are different acts. The same separation is worth imposing on an agent: an agent that reads a code, a link or an address in a file should report it, not visit it. And an agent steering your own browser is holding whatever that browser is signed into — a reason to give it the same squint you give any tool, which is the next section.

An agent can check the promise too

The four checks the upload guide teaches — pull the plug, watch the Network tab, read the Content-Security-Policy, read the code — are all things an agent can run, and they are actually easier for an agent than for a person: reading a CSP header or searching served source for fetch calls is mechanical work. If you use an agent to vet tools before trusting them, this site expects to be vetted the same way, and the offline behaviour the checks rely on has a page of its own.

The pages label their controls and publish llms.txt and Markdown copies for reading. Their source, shared modules and bundled engines are public. The build generates and minifies the served pages, so compare a build with the deployed output when auditing code. Availability of DevTools and WebRTC diagnostics depends on the agent environment.

This guide cannot make a privacy guarantee about an agent vendor. Instructions, screenshots and sometimes files may be sent by the agent even when its browser is local. Treat that data path and the tool's documented network features as part of the task, not as something browser location alone proves.

Tell us what could work better

Feedback from people and agents is welcome. Send private feedback without an account, login or email address. The same page documents a JSON endpoint for agents that can send HTTP requests.

Name the page and what you were trying to do, what you expected and what happened, or the improvement you propose. Agent and browser details help when a control or permission got in the way. Use a small synthetic example; leave out private files, credentials and conversation transcripts. Sending feedback deliberately sends the text through Cloudflare to the maintainer’s email inbox. An agent should post within the person’s authorization, or prepare a draft for them.

Declare whether you are a person or an agent; the label is self-declared and is recorded separately from the submission format. An agent using the form should still choose “An agent.” Combine related observations in one report, keep its receipt, and avoid repeated submissions. The service suppresses duplicate messages for the same page and applies network and global sending limits. Follow any retry time it returns; pending or unconfirmed delivery is not a reason to send again automatically.

Code contributions are welcome too. Fixes, clearer documentation, accessibility improvements and new tools all help. Read the contribution guide and open pull requests against dev. GitHub requires an account for a pull request; the anonymous feedback form can also take a proposed fix or a public patch link.