Skip to main content
A project chat is an AI Studio conversation locked to one project. The agent only reads and changes that project’s data, keeps its scans inside the project’s scope, and follows the instructions and reference files the project’s members set for it. Use a project chat for any engagement that has a defined scope: a pentest of one perimeter, a review of one application’s findings, a compliance check of one environment.

Start a project chat

1

Open a chat for the project

In AI Studio, click the project under Projects in the sidebar. Or open the project from Projects in the console and click Chat with agent.
2

Check the scope and mode

The project shows above the composer, and the chat is in Agent mode. Click Agent to switch back to Ask mode if you only have questions. Until you send, you can still pick another project, or Whole workspace.
3

Send your request

Type what you want done, or pick a starter prompt, and press Enter. From now on the chat is locked to the project.
The sidebar and the project selector list your active projects. Starting a chat needs agent.read; the project page also needs project.read.

What the lock means

Once you send the first message, the project is fixed for the life of the chat. It shows above the composer with its color; hover over it to read “The agent only reads and acts inside this project”. To work on another project, or on the whole workspace, start a new chat. Inside a project chat:
  • Workspace data stays in the project. Every lookup the agent makes (assets, vulnerabilities, ports, technologies, TLS certificates, statistics) and every project tool only returns and changes data of this project, whatever the agent asks for. When you ask about something outside it, the agent says so.
  • Scans stay in the project’s scope. When the project has include rules, the scope check uses the project’s include and exclude rules instead of the workspace targets. A tool run or subagent aimed outside them is refused; a host outside them named in a shell command asks for your approval, even in Auto; an excluded host is always refused.
  • The agent follows the project’s instructions and reads its files, described below.
AI Studio new chat in Agent mode with the Acme Fintech – PCI External Perimeter project selected above the composer and the worker selector set to Auto (any worker)

A project chat: the project is shown above the composer for the life of the chat

Project instructions and files

The AI agent card on the project’s Settings tab holds the context every AI Studio chat started from the project receives.
Project Settings tab with the Chat with agent button in the header, the Information card, and the AI agent card with an Instructions box using 136 of 2000 characters and a Files section showing 0 of 5 files

The AI agent card on a project's Settings tab

Write the instructions

1

Open the project's settings

Open the project and click the Settings tab.
2

Write the instructions

In Instructions, tell the agent how to work on this project: what to focus on, what to stay away from and how to report back. For example: “Only test the staging hosts and rank findings by business impact”. The counter shows how many of the 2,000 characters you used.
3

Save

Click Save under the box. The console confirms with “Instruction saved”. Clearing the box and saving removes the instructions.

Attach reference files

1

Upload a file

Under Files, click Upload and pick a .md, .txt or .csv file of up to 10 MB, such as scope notes, rules of engagement or an asset list.
2

Check the list

The file shows with its size and who added it. The counter next to Files shows how many of the five slots are used; Upload is disabled when all five are.
3

Detach a file

Click × next to a file to detach it from the project.
Files are stored in the workspace’s Files and count toward its storage. Detaching a file from the project keeps it there. Changing the instructions or files needs the project.write permission. See Projects.

How the agent uses them

At every turn of a project chat, the agent receives:
  • the project’s name, and that the chat is locked to it,
  • the instructions, marked as set by the project’s members, for it to follow,
  • the start of each attached file: up to 8,000 characters per file and 30,000 characters across all files.
When a file is longer than what it received, the agent knows its full length and reads the rest page by page when it needs it. The agent can also look up the project itself: its description, scope rules, attached files, workflows and triggers. Changes take effect at the next message, in every chat of the project, including chats already open.
Put lasting rules of engagement in the project, not in each chat: every chat of the project, and every member’s, follows them. Put preferences that apply to all your chats in your personal instruction.
In Agent mode the agent can also replace the project’s instructions itself, for example when you ask it to record a new rule. That needs your project.write permission and, in Manual mode, your approval like any other change.

Archived projects

An archived project is read-only, and so is its AI agent card: you cannot change the instructions or files until the project is reopened.
  • Archived projects are not listed under Projects in the AI Studio sidebar, or in the project selector.
  • Chat with agent on an archived project’s page still opens a chat locked to it.
  • In a chat about an archived project, the agent is told the project is archived and read-only. It can still read the project’s data and its instructions and files, but changes to the project, such as its scope, workflows or instructions, are refused until it is reopened.
See Archive, reopen and delete.

Deleting a project deletes its chats

Deleting a project also deletes every AI Studio chat locked to it, for every member, with their messages, plans and subagents. The confirmation says so: “Its workflows, their run history and its AI Studio chats go with it.” Vulnerabilities and assets stay in the workspace. Deletion cannot be undone. To keep the chats, archive the project instead.

Projects

Scope, workflows, triggers and the AI agent card

Chat

The composer, modes and conversations

Plans and approvals

The scope check and the other safety checks

Instruction

Personal instructions for all your chats