Permissions, browser, web search, code index, memory, skills, MCP, sub-agents and context.

The third group in Settings, "Agent", covers what the agent can do and how. Pages marked advanced appear only after you turn on "Show advanced settings" at the bottom of the sidebar.

Permissions

Ask before acting

Level	Behavior
Only risky actions (default)	Edits files and runs commands directly; asks before deleting, resetting or force-pushing
Every step	Asks before each action
Never	Blocks nothing; you're asked to confirm when switching to it

Other switches:

Don't ask twice for the same command
Ask before sending or deleting after reading web content — after it has read a web page, document or the screen, it asks before sending messages, committing or deleting files, so instructions hidden in content can't act on their own (off when "Never" is selected)
Strip secrets before sending — passwords, keys and tokens are replaced with <REDACTED> before reaching the model
Sign commits as Neox — adds Co-Authored-By: Neox to commit messages
OS-level sandbox — limits where commands can write (read-only / workspace / unrestricted) and whether they can use the network; safer, though some commands may not run

This page is hidden in Work mode.

Browser

The agent browses in its own separate Chrome and leaves the browser you use alone. You can bring along the sign-ins from one of your everyday Chrome profiles (synced on demand) and limit which sites it can open with allow / block lists. Hidden in Work mode.

Web Search advanced

Turn web search on or off and pick the source: built-in (no setup once signed in), Bocha, Serper (Google results) or a custom endpoint.

Code Index advanced

See the index status; build, update, rebuild from scratch or clear it; set exclusion rules. See Code index.

Memory

See and delete what Neox remembers (things you asked it to remember, conventions you set) and recent work. Each project keeps its own. See Memory.

Skills

List, search, create and import skills and see their call history; skills you already have in Claude Code / Codex can be imported in one click. See Skills.

MCP

Add and manage MCP servers (local commands or remote HTTP / SSE), see connection status and tools, or turn MCP off entirely. Called "Services" in Work mode. See MCP.

Sub-agents advanced

Sub-agent model: set per main model; by default a faster, cheaper model from the same family is picked
Auto-generate session titles
Nesting depth: how many levels sub-agents can dispatch further (0 = none)
Concurrent teams

Hidden in Work mode. See Subagents.

Context advanced

What the current context is made of, with compress-now and clear buttons
Request mode — standard (about 17–26K tokens at the start of each request) or lean (about 3K)
Auto-compress — on/off, trigger threshold, and sync or background compression

Settings · Agent