AI agents

Agent tools

Reference of the tools Jennefer agents use in the IDE: file tools, search, commands and background dev servers, collaboration, plans, web pages and MCP, with approval rules and command limits.

On this page

Tool reference
Files
Commands
Background dev servers
Team work
Web pages
MCP tools
Command limits
PDFs and images
What agents cannot do
FAQ

Agents act through a fixed set of tools. This page lists them, which need your approval, and the limits that apply. Which tools an agent actually has depends on its role and on its settings in Hub > Agents.

Tool reference#

Files#

Tool	What it does	Approval
`read_file`	Reads a file, or a range of lines. PDFs are turned into text.	No
`list_dir`	Lists the contents of a folder.	No
`search_files`	Searches the text of every file in the project, or in one folder.	No
`create_file`	Creates a file with the given content.	No
`edit_file`	Replaces exact text in a file. The text must match exactly once. Several changes can go in one edit.	No
`create_directory`	Creates a folder.	No
`delete_file`	Deletes a file.	Yes
`delete_directory`	Deletes a folder. The project folder itself can never be deleted.	Yes, every time
`move_file`	Moves or renames a file or folder.	Yes

The default roles do not have delete_file, delete_directory or move_file. You can give them to an agent in Hub > Agents if its role allows them.

Commands#

Tool	What it does	Approval
`execute_command`	Runs a command in your project folder and returns its output. With the background option it starts a long-running server instead. See below.	Yes

Commands run hidden in PowerShell, in your project folder. Their output shows in the Terminals chip under the chat box. Programs that always ask, and the checks a team run makes without asking, are listed in Approvals and safety.

Background dev servers#

Agents can start your app's dev server, for example npm run dev, in the background so you can open the app in your browser:

The agent asks with the normal card: wants to run a terminal command. Allow it? with the command.
After you approve, the server starts in your project folder and keeps running. The agent waits a few seconds for its address, such as http://localhost:3000, and tells you.
The bell shows Your app is running with the address and the command. Open the address in your browser.
The server keeps running after the run ends, and after you stop a run.
Closing the IDE stops every server its agents started.

Team work#

Tool	What it does	Approval
`ask_user`	Asks you a question as text, a choice or a masked secret.	You answer it
`ask_agent`	Asks another team member for input.	No
`handoff_to_agent`	Passes work to another team member.	No
`task_complete`	Marks the current task as done, not possible, or needing your input.	No

Agents can also write and edit plans and documents. They open from the Plans chip under the chat box.

Web pages#

Tool	What it does	Approval
`web_fetch`	Reads one web page as text, through the Engine on your computer. Agents cannot search.	No for links you wrote yourself. Other pages follow Hub > Settings > Agents and the web.

See Agents and the web for the Off, Ask and On settings and the limits.

MCP tools#

Agents can also use tools from the MCP servers you selected for the chat. Every call asks for your approval. See MCP servers.

Command limits#

Limit	Value
Longest run of one command	10 minutes
Stall detection	A command with no output for 2 minutes is stopped
Output buffer	10 MB per command
What the model sees	About 20,000 characters of the output, from the start and the end

A timeout or a stall is treated as an environment problem: the agent reports it instead of retrying forever. A single agent cannot repeat a command that just failed without changing something first. Background servers are not stopped by these limits.

PDFs and images#

Reading PDFs: read_file turns a PDF into text, so a requirements document in your project can drive a plan. Scanned PDFs without a text layer may give little or nothing.
Images: attach them to your message. They need a model marked Reads images in Hub > Models.

What agents cannot do#

Search the web. Agents read only pages you link, or pages they asked you about first. See Agents and the web. Other network access happens only through commands such as npm install or git clone, and curl and wget always ask first.
Leave the project folder. File tools are confined to your project, unless you approve a folder outside it for the run.
Make images or videos. That is Studio.
Work after you close the IDE. Runs stop when the IDE closes.

FAQ#

Can an agent install packages?

Yes, through commands like npm install after you approve them. Approve all for this run can cover one program, for example npm.

Do tools work the same with one agent and with the whole team?

Yes, within each agent's tools. Plan first changes nothing.