AI Computer Control

Beyond querying your fleet and running scripts, the Breeze AI assistant can look at and interact with a device’s own desktop: it can capture a screenshot, analyze what’s on screen against a description you give it, and — with a second person’s approval — move the mouse and type on the device directly. This page covers those three tools, plus the two remote-session tools that sit in the same part of the codebase, and the separate AI Operator preview for handing off a longer-running task.

Note

This page does not cover the AI’s broader device-management toolset — running commands and scripts, managing services, file operations, disk cleanup, security scans, and so on. Those are documented on AI Features, along with the full tier system and approval workflow this page relies on.

---

Available Tools

Tool	Description	Tier	Approval
take_screenshot	Capture a screenshot of the device screen	Tier 3	Supervised — approve in chat
analyze_screen	Capture a screenshot and analyze it against a context you describe	Tier 3	Supervised — approve in chat
computer_control	Send mouse/keyboard input and capture a screenshot after each action	Tier 3	Four-eyes — a second person approves
list_remote_sessions	List active and recent remote sessions (terminal, desktop, file transfer)	Tier 1	No
create_remote_session	Open a new remote terminal or file-transfer session to a device	Tier 3	Four-eyes — a second person approves

Caution

computer_control and create_remote_session are classified four-eyes, not supervised — the requesting technician cannot approve their own request. Breeze treats direct input control and unattended remote-session creation as surveillance-grade access, so a different user has to approve it from Approval History (or the mobile app). take_screenshot and analyze_screen stay supervised: the requester approves their own screenshot request in chat. See AI Features → Approval workflow for what four-eyes and supervised mean and who can approve each.

---

Screen capture & analysis

take_screenshot

Captures a screenshot of the device screen and returns the image for visual analysis. The device must be online.

Parameter	Description
deviceId	The device UUID (required)
monitor	Monitor index to capture (default: 0 = primary)

"Take a screenshot of DESKTOP-A1B2C3"
"Show me what's currently on the CFO's screen"

analyze_screen

Takes a screenshot and analyzes what’s visible against a context you describe — for example, “error dialogs” or “performance issues.” Useful for troubleshooting what the user sees without opening a full remote session.

Parameter	Description
deviceId	The device UUID (required)
context	What to look for or analyze on screen
monitor	Monitor index to capture (default: 0)

"Check DESKTOP-FINANCE-01's screen for any error dialogs"
"Is the user stuck on a Windows Update screen?"

---

Computer control

computer_control

Sends mouse and keyboard input to a device and returns a screenshot after each action by default. The device must be online.

Action	Requires
screenshot	—
left_click, right_click, middle_click, double_click	x, y
mouse_move	x, y
scroll	x, y, scrollDelta
key	key (optionally modifiers: ctrl, alt, shift, meta)
type	text

"Click the 'Retry' button at 640, 480 on SERVER-01"
"Type the license key into the activation dialog on DESKTOP-A1B2C3"

Note

Every computer_control invocation is four-eyes: the AI raises the proposed action, and a different user has to approve it from Approval History before it reaches the device. There is no self-approval path for this tool, even for a sole administrator — see AI Features → Approving your own request for what happens when an organization has no second eligible approver.

---

Remote session tools

Two tools give the AI limited visibility into and control over Remote Access sessions. Full session mechanics (WebRTC, TURN, WebSocket fallback, Quick Support) are documented on that page — this is only what the AI assistant itself can do.

list_remote_sessions

Lists active and recent remote sessions, filterable by status, type, and device. A non-system caller only sees their own sessions. Tier 1, read-only, no rate limit.

create_remote_session

Opens a new terminal or file_transfer session to an online device. desktop sessions are not created through this tool. Four-eyes approval, same reasoning as computer_control: an unattended session is surveillance-grade access. If the organization is not verified for remote control, the tool returns an error rather than opening a session.

"Open a terminal session to SERVER-DB-02"

---

Safety & limits

Timeouts

All three device-control tools use a 120-second timeout — the same budget as execute_command, since each does an agent round-trip and, for analyze_screen, an additional vision pass.

Tool	Timeout
take_screenshot	120s
analyze_screen	120s
computer_control	120s

Rate limits

Tool	Limit	Window
take_screenshot	10 requests	5 minutes
analyze_screen	10 requests	5 minutes
computer_control	20 requests	5 minutes
create_remote_session	10 requests	5 minutes

Permissions

Tool	Required Permission
take_screenshot	devices:execute
analyze_screen	devices:execute
computer_control	devices:execute
list_remote_sessions	devices:read
create_remote_session	devices:execute

Production allowlist

Screen contents are the most sensitive RMM output there is — credentials, customer data, anything on the display. In production, take_screenshot, analyze_screen, and computer_control additionally require an explicit entry in the MCP_EXECUTE_TOOL_ALLOWLIST environment variable, same as any other Tier 3+ tool reached through the MCP server. See MCP Server → Production allowlist.

---

AI Operator (preview)

AI Operator is a different shape of automation from the tools above: instead of a single tool call inside a chat turn, it’s a long-running task — investigate, act, verify, document — that keeps working after you close the browser, pausing only at the step that needs your approval.

Caution

AI Operator ships disabled, behind two feature flags plus the main BREEZE_AI_AGENTS_ENABLED kill switch, and only one recipe — service recovery — currently exists. It’s documented in full on AI Agents → AI Operator (preview), including the Delegate to Operator button, what the task page shows, and the service-recovery approval flow. See Environment Variables → AI Operator for the flags.

---

Troubleshooting

Tool execution times out. take_screenshot, analyze_screen, and computer_control all use a 120-second timeout. If the device is under heavy load or has a slow connection, retry once it settles.

“Device not online” error. All five tools on this page require the device to have an active WebSocket connection. Check the device’s status in the dashboard; if it shows online but commands still fail, the agent may need to be restarted.

Approval request not appearing in the UI. Tier 3 actions emit an approval_required event to the AI chat session. computer_control and create_remote_session requests don’t show a self-approve option even for the requester — a different user has to approve from Approval History or the mobile app. See AI Features → Pending approvals survive a restart.

“Tool rate limit exceeded” error. Each tool’s per-user limit resets on its own window (see Rate limits above). The error message includes the reset time.

“Not in MCP_EXECUTE_TOOL_ALLOWLIST” error. In production, take_screenshot, analyze_screen, and computer_control must be explicitly listed in MCP_EXECUTE_TOOL_ALLOWLIST to be callable through the MCP server. See MCP Server → Troubleshooting.