AI Copilot

Copilot is Studio’s context-aware AI assistant. It lives in the right-side panel. It can see your active tab, your terminal selection, attached images, your host inventory, your procedures, and your saved memories. It can open new tabs, run terminal commands with approval, call your connectors, build diagrams, and write artifacts. Copilot is designed to let you move fast when the context is clear and slow down when it isn’t. The way you express that preference is the mode you pick and the approval posture you keep.

Response modes and Autopilot

Studio separates response mode from the higher-risk autonomy toggle. The response mode menu gives you Default, Ask, and Planning. Autopilot appears below those modes as a separate permission that auto-approves tools and removes the turn limit.

Ask

Conversational and read-only. Copilot can explain output, summarize evidence, look things up, and answer questions — but it will not run commands that change state. Use this when you want a second opinion, a walk-through, or a summary.

Planning

Copilot researches first and proposes a plan before touching anything. It uses read-only tools to gather evidence, then presents a list of proposed steps for you to approve. Use this for production changes, incidents on unfamiliar devices, and multi-step diagnostics.

Default

Full tool access. Copilot acts as needed and pauses for approval when it hits something moderate, dangerous, or unclassified. Use this for routine work where you trust the context and the approval prompts are enough friction.

Autopilot

A permission toggle, not a response mode. It auto-approves tools and runs without a turn limit. Use this only in controlled environments where the target, credentials, and blast radius are unambiguous.

Autopilot skips the per-tool approval prompts. Only use it when you’ve already decided the scope is safe.

Choosing the right posture

Use the lightest posture that still lets you move:

Situation	Start with	Why
You want an explanation of selected output.	Ask	It stays conversational and read-only.
You are on an unfamiliar production device.	Planning	Copilot gathers evidence and proposes a plan before acting.
You are doing routine diagnostics on a known host.	Default	It can use tools and pause where approval matters.
You are running a bounded lab task.	Default plus Autopilot	The target and blast radius are controlled.
You are about to write a reusable runbook.	Planning	It produces a cleaner sequence and safety assumptions.

When in doubt, begin in Ask or Planning. You can move to Default once the evidence is clear. Treat Autopilot like a scoped automation setting, not a convenience switch.

What Copilot can see

The power of Copilot is the operational context around the conversation. Attach context explicitly when you want precision:

Context	What it adds
Active tab	The state of whatever tab you’re on — terminal buffer, editor contents, procedure body, dashboard data.
Terminal selection	A focused block of output. Right-click or use the selection toolbar to send it to Copilot.
Image attachments	PNG, JPEG, GIF, or WebP — for diagrams, hardware photos, whiteboard sketches, or console output you couldn’t paste.
Voice	Dictate prompts through the voice control.
Hosts	Copilot can inspect your inventory, protocols, device identity, and folders.
Procedures	Copilot can create, edit, or run procedures.
Memories	Copilot can search and save structured facts.
Connectors	Copilot can call your REST and SOAP connectors (with approval).
MCP servers	Tools from your connected MCP servers become callable.
Generated artifacts	Markdown reports, data tables, code, dashboards, diagrams, diffs — everything Copilot has produced in the conversation.

The prompt composer shows context usage so you can tell when a conversation is becoming large. Compact or start a fresh thread when old context is no longer helping. Attach the exact tab or selection when precision matters; relying on ambient workspace context is convenient, but explicit context produces better answers.

What Copilot can do

Copilot has access to the same operational surface you do, grouped by domain. Most of these are read-only by default; the ones that change state require approval.

Domain	Examples
Connectivity	Ping, traceroute, DNS lookup, ARP scan, SNMP get/walk, path MTU, interface info.
Device access	Open SSH, Telnet, serial, or RDP sessions; run commands; read output; inspect remote-session context.
Layer 2 diagnostics	LLDP and CDP discovery, passive ARP monitoring, STP observation, VLAN detection.
Security & fingerprinting	TCP banner scans, vendor and OS fingerprinting on reachable devices.
Parsing & text	Turn CLI output into tables, extract values with regex, diff configs, convert formats.
Network math	CIDR arithmetic, subnet checks, OUI lookup, ASN lookup, port-to-service lookup.
Inventory	Search, create, update, and delete hosts, subnets, IPs, diagrams, memories.
Diagramming	Create or update a network map from inventory, evidence, or a described topology.
Artifacts	Open markdown, data tables, code editors, status cards, forms, charts, and diffs inside the conversation.
Knowledge	Semantic search, knowledge-graph queries, and retrieval over your saved memories.
Session recording	List, replay, and search recorded SSH or Telnet sessions.

Some actions are inherently external: connector calls, API writes, inventory edits, and device commands can affect systems outside Studio. Copilot should show what it intends to do before it does it. Review the target, payload, and credential context before approving.

Approval posture

Studio classifies every tool call before running it. Approvals scale with risk:

Risk class	Examples	Default behavior
Read-only	`show`, `display`, `get`, ping, traceroute, inventory lookups.	Allowed in any mode.
Moderate	Entering config mode, interface changes, ACL edits, saving configuration.	Requires review unless you loosen it explicitly.
Dangerous	Reload, erase, format, force-delete, clearing routing processes.	Requires explicit approval or is blocked.
Unknown	Commands Copilot can’t classify with confidence.	Treated as review-required.

Ask mode and Planning mode are the safest places to start a new investigation. They keep Copilot in an evidence-gathering posture while you decide how to proceed.

Slash commands

Slash commands change session behavior without typing a full sentence.

Command	What it does
`/compact`	Compact the conversation history to free up context.
`/clear`	Start a fresh conversation.
`/model fast`, `/model balanced`, `/model deep`	Switch the model tier when tier aliases are available in your build. Use settings to pick the default model.
`/effort low`, `/effort medium`, `/effort high`, `/effort max`	Set the effort level for the current run.
`/thinking adaptive`, `/thinking budget`, `/thinking off`	Control how much reasoning Copilot does before replying.
`/plan`	Enter planning mode.
`/trust autonomous`, `/trust supervised`, `/trust manual`	Change the approval posture when that command is available in your build.
`/devices`	Show connected devices.
`/audit`	Show the recent action audit trail.
`/help`	List available commands.

Streaming and steering

While Copilot is working, you can watch progress in real time — thinking state, tool calls, turn count, token usage. You can steer the current turn by sending a new message, queue a steering message for the next available action, or stop the run outright when you’ve seen enough. Copilot picks up the queued message at the next natural break. Use steering aggressively during live operations. If Copilot starts reading the wrong host, assumes the wrong VRF, or takes a plan in the wrong direction, interrupt with the correction instead of waiting for a polished answer. The transcript will still show what happened.

Multi-agent work

For complex or long-running tasks, Copilot can delegate to specialist agents — a researcher for read-only context gathering, an executor for coordinated multi-step runs, focused agents for terminal operations, browser automation, network discovery, and procedure authoring. Each sub-agent inherits the same approval posture and tool restrictions you set, so a delegation never escalates beyond what the parent conversation allows. You see their progress inline in the same conversation, and the final result rolls back up as one transcript.

Good prompts

Goal	Prompt shape
Explain	”Explain the selected BGP output and list the evidence for each conclusion.”
Diagnose safely	”Use read-only checks to determine why this VLAN isn’t passing traffic.”
Build a change plan	”Plan the exact checks and commands for this maintenance window — do not execute changes.”
Use a connector	”List available endpoints on the monitoring connector, then show which one can fetch current alarms.”
Create a procedure	”Turn this successful path into a reusable procedure with arguments for hostname and interface.”
Preserve knowledge	”Save the conclusion and affected devices as a memory tagged with this incident.”

A safe incident pattern

Start in Ask and explain the symptom in plain language.
Attach the active terminal, selected output, or relevant artifact.
Move to Planning and ask for read-only checks only.
Review the proposed commands and remove anything outside the incident scope.
Switch to Default only when you are ready for Copilot to run approved tools.
Ask for a final artifact: summary, evidence, timeline, and follow-up.
Save durable facts as memories and promote the repeatable path to a procedure.

Procedures

Turn a successful conversation into a replayable runbook.

Connectors and MCP

Give Copilot access to your APIs and third-party MCP servers.

Start here

Operate devices

Work with AI

Platform

Reference

Response modes and Autopilot

Choosing the right posture

What Copilot can see

What Copilot can do

Approval posture

Slash commands

Streaming and steering

Multi-agent work

Good prompts

A safe incident pattern

Procedures

Connectors and MCP

​Response modes and Autopilot

​Choosing the right posture

​What Copilot can see

​What Copilot can do

​Approval posture

​Slash commands

​Streaming and steering

​Multi-agent work

​Good prompts

​A safe incident pattern

​Related docs

Procedures

Connectors and MCP

Response modes and Autopilot

Choosing the right posture

What Copilot can see

What Copilot can do

Approval posture

Slash commands

Streaming and steering

Multi-agent work

Good prompts

A safe incident pattern

Related docs