// PRACTICAL GUIDES

Give the agent room. Keep the boundary.

Concrete setups for the moment when “please be careful” stops being a security plan. Start with the problem that sounds like yours.

Choose a guide
A coding agent receives instructions and untrusted context, then agentjail checks tool policy and operating system access before work can run.
// EXPLORE THE BOUNDARY

Follow a tool call from request to result.

This is the security model the guides build on. Open the full map to focus nodes, trace routes, or switch views.

Open the full-screen map ↗
// PICK YOUR STARTING POINT

What are you trying to make safer?

You do not need to read these in order. Each guide starts with a real setup and ends with something you can test.

AUTONOMY 01

I want fewer permission prompts

Claude Code: safely skip permission prompts

How agentjail enforces a hard boundary when you run an agent with permission prompts disabled.

Read the guide
CLAUDE CODE 02

I need to contain Claude Code

Claude Code sandbox

Set up Claude Code with tool policy and OS sandboxing, then test filesystem, network, and MCP restrictions.

Read the guide
MCP 03

I have MCP servers to review

MCP security guide

Inventory MCP servers, scope individual tools and credentials, and test policy before allowing agent access.

Read the guide
SECURITY MODEL 04

I need the full security model

AI agent security guide

Map a coding agent's permissions, enforce boundaries outside the model, and test the controls that limit mistakes.

Read the guide

Looking for flags, config fields, or policy syntax?

Open the full documentation →