What the AI Can Do

The assistant is not a chat box with your code pasted into it. It has 61 tools it can call directly against your project, your running server and your database. This page is the map of that surface.

Every tool carries one of three risk levels, and the risk level decides whether it runs silently or stops to ask you:

Risk	Count	Behaviour
read	41	Runs silently. Nothing changes.
write	14	Asks first. Changes files, config or data.
execute	6	Asks first. Runs something outside Boltise.

See Approvals & Safety for how the prompts work and what can never be approved at all.

Files

Reading, searching and editing your project.

Tool	Risk	What it does
read_file	read	Reads a text file, with paging for long ones
find_files	read	Finds files by name
grep_project	read	Searches file contents
list_dir	read	Lists one directory
edit_file	write	Changes part of a file
write_file	write	Writes a whole file
copy_file / move_file / delete_file	write	Moves files around
revert_file	write	Puts one file back to how it was when the conversation started

read_file returns the raw bytes with no line numbers, because every edit has to match what is actually on disk.

Lua and manifests

Boltise parses Lua with a real parser, not regular expressions, so these answers are structural rather than textual.

Tool	Risk	What it does
validate_lua	read	Checks a snippet or file parses
security_audit	read	Flags unsafe patterns in Lua
check_fxmanifest	read	Validates an fxmanifest.lua
lookup_native	read	Looks up a FiveM/RedM native and its real signature
joaat_hash	read	Computes a native hash
lsp_diagnostics / lsp_hover	read	Editor-grade diagnostics and type info
migrate_lua	write	Rewrites ESX-style Lua toward another framework
generate_luacats_stubs	write	Generates native type stubs for autocomplete
scaffold_resource	write	Creates a new resource skeleton

Server config

Tool	Risk	What it does
validate_cfg	read	Checks server.cfg parses
edit_cfg	write	Changes a convar or ensure line

server.cfg gets its own editing tool rather than going through edit_file, because a broken line there stops the server from booting at all.

The live server

These only work against a server Boltise itself started.

Tool	Risk	What it does
server_status	read	Is it up, and for how long
list_resources / list_players	read	What is running, who is on
read_server_console	read	Recent console output
get_server_events / get_resource_metrics	read	Live event and metric feed
diagnose_resource	read	Why one resource is misbehaving
restart_resource	execute	Restarts one resource
server_command	execute	Runs a console command
trigger_event	execute	Fires a server event
execute_lua_live	execute	Runs Lua on the running server

execute_lua_live is exactly as powerful as it sounds. It always asks.

Performance

Tool	Risk	What it does
capture_profiler_snapshot	execute	Records a profiler trace and parses it
start_capture / wait_for_capture / cancel_capture	read	The same capture, driven step by step
get_slow_queries	read	Slowest database queries

See Resource Monitor for what a capture contains.

Database

Tool	Risk	What it does
query_database	read	Runs a read-only query
execute_database_write	write	Runs a statement that changes data

Reads and writes are deliberately separate tools, so a SELECT never needs your approval and an UPDATE always does.

Code Graph and NUI

Tool	Risk	What it does
read_code_graph	read	Reads the parsed structure of a resource
open_code_graph	read	Opens the graph panel on a symbol
nui_preview_open / _close / _refresh / _restart	read	Drives the NUI preview
nui_preview_screenshot	read	Captures what the NUI looks like
nui_preview_send_message	read	Sends a message into the NUI
render_canvas	read	Renders a diagram or chart into the chat

Memory and planning

Tool	Risk	What it does
memory_search / memory_list	read	Looks things up in long-term memory
memory_save	write	Remembers something
save_plan / update_plan	write	Keeps a visible task list for long jobs
get_chat_history	read	Re-reads earlier in the conversation
load_skill	read	Loads a task-specific instruction set

memory_save is the one write that does not show an approval card — an owner decision, on the grounds that saving a memory is not risky. It is still recorded in the audit trail. See Project Memory.

The outside world

Tool	Risk	What it does
web_search	read	Searches the web
web_fetch	read	Reads a specific page
run_shell_command	execute	Runs a command in your terminal
ask_user	read	Stops and asks you a question

run_shell_command is the most heavily guarded tool in the app. Read Approvals & Safety before you loosen anything around it.

Note

Tools are gated by what is actually available. With no server running, the live server tools are not offered at all rather than failing when called; the same goes for the database tools with no connection.