Quick Start

Get RADIUS running in 5 minutes with any agent framework — or no framework at all.

Using OpenClaw? Skip to the OpenClaw Quick Start for the full golden-path walkthrough including Telegram approvals.

Prerequisites

• Node.js >= 20
• npm (comes with Node)

Step 1 — Install

npm install agentradius

This adds the agentradius package to your project and makes the npx agentradius CLI available.

Step 2 — Initialize

npx agentradius init --profile standard

This scaffolds a default radius.yaml in your project root and creates a .radius/ directory with wiring files.

No --framework flag means generic mode — RADIUS will run as a standalone pipeline you can call via HTTP, MCP, or library import.

Profile options

Start with standard for development. Switch to local before production.

Step 3 — Tour the config

Open the generated radius.yaml. Here’s what each section does:

global:
  profile: standard             # Security posture
  workspace: "${CWD}"           # Agent's workspace root
  defaultAction: deny           # Unlisted tools are denied

modules:                        # Pipeline order matters
  - kill_switch                 # Emergency stop
  - skill_scanner               # Detects prompt injection in skills
  - tool_policy                 # Allow/deny rules per tool name
  - fs_guard                    # Filesystem path constraints
  - command_guard               # Shell command pattern blocker
  - exec_sandbox                # bwrap namespace isolation
  - output_dlp                  # Secret detection & redaction
  - rate_budget                 # Sliding window rate limit
  - audit                       # Append-only event log

moduleConfig:
  fs_guard:
    allowedPaths:
      - "${workspace}"
      - "/tmp"
    blockedPaths:
      - "~/.ssh"
      - "~/.aws"
      - "/etc"
  command_guard:
    denyPatterns:
      - "(^|\\s)sudo\\s"
      - "rm\\s+-rf\\s+/"
  rate_budget:
    windowSec: 60
    maxCallsPerWindow: 60

The modules array defines the pipeline order — every event passes through each module from top to bottom. The first deny wins.

Template variables like ${workspace}, ${CWD}, ${HOME} are resolved at load time. You can also use environment variables: ${TELEGRAM_BOT_TOKEN}.

Step 4 — Health check

npx agentradius doctor

The doctor command verifies:

• Config valid — YAML parses, all referenced modules exist
• Modules loaded — each module initializes without error
• Workspace readable — the workspace path exists and is accessible
• bwrap available — (if exec_sandbox is enabled) checks for bubblewrap binary
• Audit writable — audit sink is reachable (file path, stdout, webhook)

All green? You’re ready.

Step 5 — Test your defenses

npx agentradius pentest

The pentest command runs a battery of simulated attacks against your config:

• Path traversal — attempts to read ~/.ssh/id_rsa
• Destructive command — attempts rm -rf /
• Privilege escalation — attempts sudo
• Secret exfiltration — outputs a fake AWS key
• Network exfiltration — attempts curl to an external domain
• Rate flood — sends rapid requests to trigger rate limiting
• Prompt injection — sends a skill with hidden instructions

Each scenario reports BLOCKED or ALLOWED. With the standard profile, all critical scenarios should be blocked.

Next steps

• OpenClaw Quick Start — framework-specific setup with hook wiring and Telegram approvals
• Configuration Reference — full annotated config, all module options, profile comparison
• Modules — per-module behavior, key options, and ordering strategy
• Features — capability map and profile selection guidance
• What’s New — latest version highlights and upgrade notes