Tools reference

The agent acts through a set of built-in tools. Which tools are available depends on the surface (CLI vs. VS Code vs. the desktop app), but the core file / shell / search set is shared everywhere. This page is the reference — for how the agent uses them in a coding loop, see the product pages.

Files

Tool	What it does
read_file	Read a file with line numbers; supports offset + limit for large files.
write_file	Create or fully overwrite a file.
edit_file	Surgical find-and-replace (unique old_string → new_string).
multi_edit	Several edits to one file, applied atomically.
delete_file	Delete a file.
move_file / copy_file	Move or copy a file.
create_directory	Create a directory (and parents).
list_directory	List a directory's contents.

Search & navigation

Tool	What it does
search_files	Grep file contents, or find files by name/glob.
glob	Find files matching a glob pattern.
get_symbols	Outline a file's functions/classes/types.
get_workspace_symbols	Search for a symbol across the workspace.
go_to_definition	Jump to a symbol's definition (VS Code).
find_references	Find all usages of a symbol (VS Code).
get_hover	Read a symbol's type + docs (VS Code).
get_diagnostics	Read the editor's live errors/warnings (VS Code).

Shell & commands

Tool	What it does
bash	Run a shell command — the main execution tool (also does navigation/linting).
bash_output / kill_shell	Poll or stop a background shell started with bash.

Task tracking & memory

Tool	What it does
todo_write / todo_read	Maintain the agent's task list.
task	Delegate a focused piece of work to a sub-agent.
memory_read / memory_write	Persistent memory across sessions.
use_skill	Load a skill playbook.

Web & data

Tool	What it does
web_search	Search the web for current information.
fetch_url	Fetch a URL (HTML stripped to text).

Documents, media & notebooks

Tool	What it does
write_docx / write_xlsx / write_pptx	Generate real Word / Excel / PowerPoint files.
generate_image / stock_photo	Generate AI art, or fetch a stock photo.
notebook_read / notebook_edit	Read and edit Jupyter notebooks.

Safety on the tool set

A few built-in guardrails shape how these tools behave regardless of the surface:

• Destructive commands require explicit confirmation.
• Silent code loss (elided edits) is refused.
• Test-integrity is enforced — weakening a test is flagged.

See Permissions for the allow/ask/deny rules that control which of these run automatically.

Next steps

• Nexrall Code for VS Code — the tools in the editor surface (semantic navigation, live diagnostics).
• Nexrall CLI — the slash commands and flags.
• Agent SDK — executeTool, the programmatic executor for these tools.