Connected tools
Let the agent work where the work is.
A browser inside Prism. Access to your files. Computer and Android phone control. An isolated server sandbox for code. Give the agent the right environment for each task, with explicit access and clear boundaries.
The connection is yours. So is the boundary.
Choose the environment, grant the access, and define where the agent should stop.
01 / Connection console
Browse. Open files. Use devices. Run code.
A browser inside Prism, your computer and Android phone, an isolated server—and your own tools.
Connection map
Only the access you grant
Choose a capability to see its own visual, setup, and limits. Illustrations do not operate your devices.
Your agent and you. The same browser.
- 01Agent browses
- 02You watch or take control
- 03Continue with the result
Server view shown. Desktop tabs stay on your computer; Android has its own on-device browser.
Inside the Prism workspace
The browser is part of your workspace.
Open pages in Prism, keep several tabs, and let the agent navigate, fill fields, and work with rendered websites. See the same web task—not just a summary of it.
What you enable
The server browser requires an active paid plan. The inline browser runs in the Prism desktop app.
- Private server browser
- Watch the agent work and take exclusive control when you need to step in.
- Inline desktop tabs
- Browse directly before, during, or after agent activity. Tabs and site sessions stay in the desktop browser.
- Short demonstrations
- Ask the agent to record the active tab as a project file: up to 30 seconds and 10 MB, without microphone or desktop capture.
Where the boundary stays
Server-browser takeover pauses agent browser actions until control returns. A browser connection does not authorize bypassing a website’s restrictions.
02 / Device boundaries
A connected device is not a blanket permission.
Your computers and phones have their own roots in Files. Choose the right device, keep it online, and grant only the access the task needs.
Three separate permission gates
Android control
- You approve
- Consent + accessibility
- Access granted
- Screen and input
Android files
- You approve
- Separate file approval
- Access granted
- Shared-storage files: read, write, copy, move, delete
Speech recording
- You approve
- Opt-in + microphone
- Access granted
- Speech clips
Outside the Android boundary: protected system files, app-private storage, and terminals. Recording permission does not grant device control.
Recording is a separate decision.
Optional background speech recording is available on Windows, Linux, and Android. It is not part of granting device control.
You opt in
Give recording consent and microphone permission. Start, stop, or revoke consent in the supported controls.
Speech clips become files
Speech detection runs on the device. Confirmed speech clips are uploaded; non-speech audio is discarded.
You ask for processing
Transcription runs on first read or search. Upload alone does not trigger a summary, indexing, or memory notes.
Illustrative browser-to-project workflow
From a web task to material you can review.
“Use the connected browser to collect the documents I select, save them to the project, and prepare a comparison. Stop before submitting anything. Let me inspect the sources and take over the browser if needed.”
Name the task and the limit
Identify the documents and browser to use. Say explicitly that the agent must stop before any submission.
Collect and compare
Let the agent open the selected sources, save the relevant documents to the project, and prepare a comparison.
This is your review point.
Inspect the sources and draft. In the server browser, take exclusive control to make your own checks; agent browser actions remain paused until you return control.
Keep the work, not just the answer
- Selected source documents
- A comparison saved to the project
- An optional short tab demonstration
Check access before you connect
- Device work requires the matching desktop or Android app, an online device, and explicit consent. Recording and Android file access have separate controls.
- The server sandbox only needs a signed-in account; its browser requires an active paid plan. Store durable server results in /server/workspace. Device capabilities vary; access does not authorize bypassing website restrictions.
- New MCP connections apply to new chats. Remote MCP servers require a supported endpoint; API keys and credentials should be treated as secrets.