How to Migrate from OpenClaw to NemoClaw: OpenShell Policy Configuration Guide (2026)

Raw OpenClaw gives AI agents unrestricted access to your file system, credentials, and network by default. NemoClaw changes that model by inserting OpenShell - a policy-based runtime - between your agents and everything they touch. Getting OpenShell policy configuration right is the migration step that determines whether NemoClaw actually protects you or just adds a layer of false confidence. This guide covers the full migration procedure and deep-dives into OpenShell policy configuration for engineers moving from a raw OpenClaw install.

What Changes: OpenClaw vs NemoClaw

Understanding what NemoClaw adds at a conceptual level prevents misconfiguration.

The core mental model shift: OpenClaw trusts agents. NemoClaw trusts policies and verifies agents against them on every action.

Pre-Migration Checklist

Complete every item before running a single install command.

Migration Steps

openclaw security audit --deep

openclaw auth revoke --all

curl -fsSL https://nvidia.com/nemoclaw.sh | bash

nemoclaw --version

nemoclaw onboard --import-from=openclaw

nemoclaw onboard

openshell policy init --deny-all

# PLACEHOLDER: confirm log path with your NemoClaw install docs
tail -f /var/log/openshell/agent.log

nemoclaw config set router.mode=local-first

logging:
  level: verbose
  destination: "/var/log/openshell/agent.log"

# PLACEHOLDER: confirm service name with your OS and OpenClaw version
systemctl disable openclaw-gateway

Post-Migration Validation

Smoke tests

1. Confirm OpenShell is active:

nemoclaw status

Expected output includes OpenShell: active and your current policy file path.

2. Confirm policy is enforcing: Attempt an action your policy explicitly denies. The action should be blocked and logged.

3. Confirm audit logging is active:

tail -n 20 /var/log/openshell/agent.log

Success signals

• All existing automation runs without unexpected policy violations
• OpenShell log shows agent actions being evaluated and allowed/denied per policy
• No trusted_proxies_missing or insecure_auth warnings in openclaw security audit --deep output
• Old credentials are confirmed revoked in target services
• nemoclaw status shows all components healthy

OpenShell Policy Configuration Guide

This is the section most migration guides skip. Getting policy configuration wrong either leaves you exposed (too permissive) or breaks your workflows (too restrictive).

The policy model

OpenShell policies are YAML files that define what an agent is permitted to do. Every agent action is evaluated against the active policy before execution. If the action is not explicitly allowed, it is denied and logged.

Policies are evaluated in this order:

1. Global deny rules (applied first, override everything)
2. Capability-specific allow rules
3. Default posture (deny-all if you initialized with --deny-all)

Authentication policy

# openshell.yaml - Authentication Policy
version: "2026-03"
agent_id: "your-agent-name"  # Replace with your agent identifier

authentication:
  require_signed_requests: true
  allowed_identity_providers:
    - "local"                  # Local machine identity
    # - "your-idp-endpoint"   # PLACEHOLDER: add your IdP if applicable
  session_timeout_minutes: 60
  revoke_on_policy_change: true  # Forces re-auth when policy is updated

Network and egress policy

network:
  mode: "allowlist"
  allow_domains:
    - "api.github.com"         # Source control
    - "pypi.org"               # Python packages
    - "registry.npmjs.org"     # Node packages
    # Add your specific API endpoints here
  deny_patterns:
    - "*.amazonaws.com"        # Block direct AWS calls unless explicitly needed
    - "*.googleusercontent.com"
  deny_all_else: true

  local_routing:
    force_local_on_sensitive: true
    sensitivity_keywords:
      - "password"
      - "api_key"
      - "secret"
      - "token"
      - "credential"

Filesystem and path policy

capabilities:
  file_system:
    allow:
      - "/home/your-username/projects/your-project/**"  # PLACEHOLDER
      - "/tmp/agent-workspace"
    deny:
      - "**/.*"              # Block ALL hidden files (.env, .git, .ssh, .aws)
      - "/home/**/.ssh/**"   # Explicit SSH key block
      - "/etc/passwd"
      - "/etc/shadow"
      - "**/node_modules/**"
    write_policy: "explicit"  # Read-only by default
    write_allow:
      - "/tmp/agent-workspace"
      - "/home/your-username/projects/your-project/output/**"  # PLACEHOLDER

Logging and auditing policy

logging:
  level: verbose          # Use verbose for first 2 weeks, then standard
  destination: "/var/log/openshell/agent.log"
  include_denied_actions: true
  include_allowed_actions: true
  rotate_policy:
    max_size_mb: 100
    keep_days: 30
  alert_on:
    - "credential_access_attempt"
    - "policy_violation_repeated"  # 3+ violations in 5 minutes
    - "write_outside_allowed_paths"

OpenClaw settings mapped to OpenShell policy patterns

NemoClaw Hardware Requirements

No NVIDIA GPU is required to install NemoClaw or run OpenShell policy enforcement. GPU hardware is only needed for local Nemotron routing - the feature that keeps sensitive workloads off cloud servers.

CUDA requirement: Local Nemotron routing requires CUDA 12.x or higher. Verify with nvidia-smi before enabling router.mode=local-first.

Password Manager Protection with OpenShell

Many engineers running OpenClaw or NemoClaw also use browser-based password managers like Bitwarden, 1Password, or Dashlane. OpenShell policies can explicitly protect these tools from agent access.

Why this matters

Password manager browser extensions store session tokens and vault unlock keys in browser profile directories. Without an explicit deny rule, an AI agent running on your machine can read these files the same way it can read any other file in scope.

OpenShell policy for password manager protection

# openshell.yaml - Password Manager Protection
capabilities:
  file_system:
    deny:
      # Mac - Chrome and Firefox profile directories
      - "**/Library/Application Support/Google/Chrome/Default/Local Extension Settings/**"
      - "**/Library/Application Support/Mozilla/Firefox/Profiles/**"
      # Linux - Chrome and Firefox profile directories
      - "**/.config/google-chrome/Default/Local Extension Settings/**"
      - "**/.mozilla/firefox/**"
      # Windows - PLACEHOLDER: adjust for your username
      - "**/AppData/Local/Google/Chrome/User Data/Default/**"
      - "**/AppData/Roaming/Mozilla/Firefox/Profiles/**"
      # Common password manager local vault files
      - "**/.local/share/1password/**"
      - "**/1Password/**"
      - "**/.bitwarden/**"

Password manager comparison: agent coexistence

Bitwarden stores vault data in an encrypted local SQLite database. The encryption is strong but the database file is accessible to any process with your OS-user permissions. The deny rules above block agent access to that path.

1Password uses a similar local vault model with additional memory protection on the unlock key. Their Unified Access feature (2026) adds an API layer for managing secrets in agentic contexts - relevant if you are building agents that need to retrieve credentials programmatically rather than reading files directly.

Dashlane operates primarily cloud-side with local caching. Less local file exposure but vault data transits cloud infrastructure by default.

For any of these: deny agent filesystem access to browser profile directories via OpenShell policy, rotate any credentials in scope before NemoClaw was installed, and use a password manager's official API or CLI for any legitimate agent credential retrieval rather than direct file reads.

Password Length Requirements After Migration

NemoClaw migration surfaces credential hygiene issues that often require updated password standards across your organization. Use this table when generating new credentials during the rotation step.

Why length matters more than complexity

The most common mistake during credential rotation is generating passwords that meet complexity rules but are too short. A 12-character password with full complexity takes approximately 3,000 years to crack with 2026 GPU hardware. A 16-character password takes 438 million years. The difference between 12 and 16 characters is not marginal - it is several orders of magnitude.

Use a cryptographically secure generator for every new credential. See the Password Length 2026 Guide for the full NIST breakdown, or generate credentials directly with the 16-character password generator for privileged accounts and the 64-character generator for API keys. For office and guest SSID credential rotation, use our WiFi password generator.

Troubleshooting

Final Migration Checklist

Related Reading

• OpenClaw Security Audit Checklist: per-warning remediation for pre-migration audit findings
• How to Uninstall OpenClaw Completely: full removal guide for post-migration cleanup
• NemoClaw vs OpenClaw: Security Architecture Comparison: conceptual overview of what NemoClaw adds
• Best Password Managers 2026: evaluated on security architecture and agent coexistence
• Password Length 2026 Guide: NIST SP 800-63B requirements for credential rotation
• 16-Character Password Generator: CISA recommended length for privileged accounts

DevSecOps engineer, M.S. Information Systems, KCSA certified. 10+ years securing enterprise systems. This guide reflects publicly available NemoClaw documentation as of March 2026. Verify commands against current official documentation before running in production.

T.O. Mercer | SafePasswordGenerator.net