Skip to main content
Every AI Studio conversation starts in the composer: the box where you type, with the controls that decide what the agent may do. This page covers each control, the difference between Ask mode and Agent mode, the starter prompts, a complete agent task from start to finish, and how to manage your conversations. Before you start, connect a model. Until you do, AI Studio shows Connect an AI Provider with a Connect button instead of the composer. See Connect a model.

Start a chat

1

Open a new chat

Click AI Studio in the console sidebar, then New Chat in the AI Studio sidebar.
2

Choose the scope

The project selector above the composer reads Whole workspace. Leave it, or pick a project to lock the chat to that project. See Project chats.
3

Choose the mode

A new chat starts in Ask mode. Click Agent to switch to Agent mode when the task needs tools run on a worker. The button turns orange and the console confirms with “Agent mode enabled”.
4

Pick a worker (Agent mode)

In Agent mode a worker selector appears next to the project selector. Leave it on Auto (any worker) or pick a worker.
5

Check the model, reasoning effort and approval mode

The model and reasoning effort sit at the bottom left of the composer, the approval mode next to the send button. The defaults are usually what you want.
6

Send your message

Type your request, or pick a starter prompt below the composer, then press Enter.
AI Studio new chat with a project selected above the composer, the worker selector set to Auto (any worker), the model and reasoning effort on the left, the Agent button active, approval mode set to Manual, and the External recon, Ports & services and Web app pentest prompt groups

A new chat in Agent mode, scoped to a project, with the starter prompt groups below the composer

The chat opens as soon as you send. Conversations are private: no other workspace member can see them.

The composer

Project

The project selector lists your active projects. A chat about a project only reads and changes that project’s data, and follows the project’s instructions and files. See Project chats.

Worker

In Agent mode the agent runs commands on a worker. With Auto (any worker), OASM uses a connected worker of your workspace, or a global one if none is connected, and the chat keeps using the worker it ran on first while that worker stays online.
  • Pick the worker before you send the first message: the chat keeps the worker it starts with.
  • The list only shows workers that are online right now, and refreshes each time you open it.
  • If the worker you picked goes offline, the selector switches back to Auto and the console says “The selected worker went offline, switched to Auto”.
  • Agent mode needs at least one online worker. With none, the Agent button opens the connect-worker dialog instead of switching mode. If you send an Agent-mode message while no worker is online, the same dialog opens and your message stays in the composer; once the new worker joins, the message is sent on it (or on Auto if the worker is not ready within 10 seconds).

Model and reasoning effort

The model selector shows the model of the connection the chat uses: your default connection in a new chat. Picking a model also saves it as that connection’s model, and the console confirms with “Model updated”. With no connection at all, the selector reads Connect LLM and opens the provider settings. The reasoning effort selector starts at Default: your settings default, else the provider’s. It lists the levels the model publishes, from None, Minimal, Low, Medium, High, Extra high and Max; a model that publishes none gets Low, Medium and High. The effort you pick is saved on the connection, next to its model. You can change the model and the effort between messages of the same chat. To change the defaults, see Models.

Approval mode

The approval selector shows the current mode; its tooltip reads “Approval: Manual” or “Approval: Auto”. A new chat starts in your Default Approval Mode from Models. Without one, it starts in the mode you last picked in this browser, or Manual the first time. See Plans and approvals for what each mode asks and the checks that run in both.

Send and stop

Press Enter or click the send button. Shift+Enter adds a new line. The send button stays disabled while the box is empty. While the agent works, the text box is disabled and reads “Waiting for response…”, and the send button turns into Stop. See Stop a run.

Ask mode and Agent mode

Ask mode

Ask mode answers questions about your workspace. The agent looks up what it needs, such as assets, targets, vulnerabilities, issues, ports, technologies, TLS certificates, statistics, workers, jobs, installed tools and projects, and answers from what it finds. For general security questions it answers from its own knowledge, and fetches a public page such as a CVE record when it needs current facts. Ask mode cannot run commands, scans or tools. When you ask for that, the agent says the task needs Agent mode and still answers whatever it can from workspace data.
Fetching a web page is not treated as read-only, so in Manual approval mode the agent asks before it fetches one, even in Ask mode. Workspace lookups never ask.

Agent mode

Agent mode does everything Ask mode does, and can also act:
  • Run commands and installed tool connectors on the chat’s worker, and drive a headless browser there.
  • Create and change projects, add targets, install tools and run project workflows.
  • Record a finding it confirmed as a proposal for review (see Findings the agent proposes).
  • Hand focused jobs to subagents that run in the background, several at once (see Side panel).
Agent mode works from a plan. When you send a task, OASM first pins a goal for it, then the agent writes the plan: an ordered list of the outcomes it will deliver. Every tool call, including a read-only lookup, waits until you approve the plan and the agent starts its first step. A call made before that is refused and shows Needs plan in the chat. A reply that needs no tool, such as a clarifying question, needs no plan. See Plans and approvals.
For a quick lookup, use Ask mode: it answers without a plan to approve.
The mode is fixed once the chat starts. To switch, start a new chat.

Starter prompts

Below the composer of a new chat, three topic chips offer ready-made prompts. Click a chip to list its prompts; hover over one (or focus it with the keyboard) to preview it as the composer’s placeholder. Click ×, press Esc or click outside the list to go back to the chips. Each group holds five prompts. Some examples:
  • Overview: “Summarize the security posture of this workspace”, “What are the top 5 risks I should fix this week?”
  • Attack surface: “Which open ports are most common, and which ones look risky?”, “Find assets that expose admin panels or login pages”
  • Vulnerabilities: “Show my critical and high vulnerabilities and how to fix them”, “Which vulnerabilities have a known CVE?”
Agent-mode prompts run real tools against real hosts, which is why they wait for you to add the target and send.

Example: run an agent task

This walk-through scans one host for exposed services, approving each command as it comes.
1

Start an Agent-mode chat

Click New Chat, pick the project the host belongs to in the project selector, and click Agent. Leave the worker on Auto (any worker) and the approval mode on Manual.
2

Pick a prompt and add the target

Click Ports & services, then “Scan my public host for open ports, then fingerprint the services on what is open and report what is unexpectedly exposed to the internet”. Type the host after the prompt, for example app.example.com, and press Enter.
3

Review the plan

The goal appears first, then the agent’s plan. The Run this plan? prompt takes the place of the composer. Read the steps, then choose Yes, but approve each command (press 2). To change the plan, type what to change in Tell the agent how to change the plan and press Enter; the agent sends a revised plan.
4

Approve the commands

As the agent works, each new command shows Run this command? with the exact command line. Press 1 (Yes) to run it, or type what to do instead and press Enter.
5

Follow the run

Tool calls appear in the chat as they run. Click Open side panel at the top right to watch the plan’s progress, the subagents the agent dispatched and every tool call. See Side panel.
6

Read the answer

The final answer opens with whether the goal was met (fully, partly or not), then the findings with their evidence and the next steps worth doing. Findings the agent confirmed may also wait for you under Vulnerabilities → Pending review (AI).
7

Follow up

Ask a follow-up in the same chat, for example “Check the TLS configuration of the services you found”. New work that needs tools gets a new plan to approve.

Answers

The answer streams into the conversation:
  • Thinking: while the model reasons, a Thinking label shows; once done it reads, for example, “Thought for 4 seconds”. Expand it to read the reasoning.
  • Tool calls appear inline as they run. Two or more calls in a row fold into one group labeled with the latest call and a count, such as Transition Step with a 46 tool calls badge; expand it to see each call. A remote command shows its output live.
  • The plan shows in its own card with its steps and a status: Planning while it waits for your approval, Plan ready once approved.
  • Copy: under a finished answer, the copy button copies its text. Older answers show it when you hover over them.
Scroll up to load older messages; the arrow button at the bottom jumps back to the latest one.
A completed Agent mode answer showing the approved four-step plan marked Plan ready, a Transition Step group with 46 tool calls, and a summary opening with Goal: met fully, next to the side panel on its Stats tab

A finished Agent-mode answer, with the approved plan and grouped tool calls above it

If the connection drops or the provider returns a temporary error, the chat retries on its own and shows “Retrying… (attempt 1 of 3)”. If it still fails, a Streaming error box shows the message with Retry, which asks for the answer again.

Follow-ups

When an answer is finished, send your next message in the same chat. The agent keeps the conversation’s history, project, mode and worker.
  • You can change the model, the reasoning effort and the approval mode between messages.
  • In Agent mode, once every step of a plan is done, new work that needs tools starts with a new plan for you to approve.
  • Long conversations are summarized automatically so they keep fitting in the model’s context window. The Stats tab of the side panel shows how much of it is used.

Runs keep going on the server

A run belongs to the server, not to your browser tab. If you close the tab, reload, open another conversation or lose your connection, the run keeps going. Open the conversation again to pick up the live answer, or read the result once it has finished. A spinner next to a chat in the Chats list shows it is still running. An approval prompt waits for you too: open the conversation again within five minutes and it is shown again. After five minutes it counts as rejected.

Stop a run

Click Stop in the composer. Stopping:
  • ends the run on the server,
  • cancels the conversation’s queued and running subagents,
  • puts the plan step that was running back to pending, so your next message can pick the plan up again.
Closing the tab never stops a run; only Stop does. Stopping needs the agent.write permission.

Conversations

The Chats list in the AI Studio sidebar shows your conversations, the most recently active first. Click one to open it, and Load more at the bottom for older ones.

Titles

A new chat is titled with your first message. After the first answer, OASM replaces it with a short title generated from the question and the answer. A chat you renamed keeps your title. Until a title exists, the chat reads New conversation. The chat’s header shows its title and how long ago it started.

Rename or delete a chat

1

Open the chat's menu

Hover over the chat in the Chats list and click … (Conversation actions).
2

Rename or delete

  • Rename turns the title into a text box. Press Enter or click away to save, Esc to cancel.
  • Delete asks you to confirm. Deleting the open chat takes you to a new chat.
To delete every conversation, click … next to the Chats heading and choose Delete all. Deletion cannot be undone. Renaming and deleting need the agent.write permission. Deleting a project also deletes its chats.

Keyboard shortcuts

Open AI Studio from elsewhere in the console

Run AI Pentest, AI gen and Chat with agent are only shown to members with the agent.read permission.
The Analyze button on a vulnerability’s AI Analysis tab is not an AI Studio chat: it writes a report on the finding with your connected model, right on the vulnerability page. See AI analysis.

Findings the agent proposes

In Agent mode, the agent can record a finding it confirmed with evidence: the requests that reproduce it, the responses that prove it and a baseline to compare against. The finding does not appear among your vulnerabilities right away. It becomes a proposal under Vulnerabilities → Pending review (AI), where a reviewer approves, rejects or merges it. The review page links back to the chat with Open conversation. The agent can only propose findings when you have the vulnerability.write permission. See AI proposals and review.

Permissions

What the agent may do with your workspace data follows your own permissions: a tool that needs a permission you lack is refused before anything runs. See Permissions.

Plans and approvals

Approve plans and commands, and the checks that run in Auto mode too

Side panel

Follow the plan, subagents, browser and tool calls of a run

Project chats

Lock a chat to a project and give the agent instructions and files

Models

Default model, reasoning effort and approval mode