Skip to content

Collen-Roller/GovClaw

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

389 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

GovClaw: Gov-Ready AI Agent Sandboxing

License Security Policy Project Status Fork

GovClaw is a government-focused fork of NVIDIA NemoClaw that adds AWS Bedrock support, Mattermost integration, and other features for defense and government environments.

It runs OpenClaw agents inside NVIDIA OpenShell sandboxes with policy-enforced network egress, filesystem isolation, and managed inference routing.

Alpha software — GovClaw inherits NemoClaw's alpha status. Interfaces, APIs, and behavior may change without notice.


What GovClaw Adds

Feature Description
AWS Bedrock First-class inference provider with API key auth, dynamic model discovery, and native bedrock-converse-stream support.
Mattermost Policy preset for DoD/gov messaging (*.cloud.mattermost.com).
Dynamic model list Fetches available models from Bedrock's ListFoundationModels API based on your region and access.
Policy fix Typing preset names directly at the policy prompt now works (e.g. "discord" applies discord, not the defaults).

All upstream NemoClaw providers remain available: NVIDIA Endpoints, OpenAI, Anthropic, Google Gemini, and Local Ollama.


Quick Start

Prerequisites

  • Linux (Ubuntu 22.04+) or macOS with Docker
  • Node.js 20+ and npm 10+
  • Python 3 (for LiteLLM proxy, optional)
  • Docker running
  • OpenShell installed

For AWS Bedrock: a Bedrock API key (long-term recommended) and model access enabled for the models you want to use.

Install

# Install OpenShell
curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh

# Clone GovClaw
git clone https://github.com/Collen-Roller/GovClaw.git
cd GovClaw

# Run the GovClaw installer
./govclaw-install.sh

The installer guides you through:

  1. Node.js — installs via nvm if missing
  2. GovClaw CLI — links the govclaw command
  3. LiteLLM Proxy — installs in a Python venv (for Bedrock)
  4. Onboarding — configure inference, create sandbox, apply policies

Using Bedrock

When the onboarding wizard shows inference options, select AWS Bedrock:

Inference options:
  1) NVIDIA Endpoints (recommended)
  2) OpenAI
  3) Other OpenAI-compatible endpoint
  4) Anthropic
  5) Other Anthropic-compatible endpoint
  6) Google Gemini
  7) AWS Bedrock (Amazon Web Services)
  8) Local Ollama (localhost:11434)

Choose [1]: 7

You'll be prompted for:

  • AWS region (auto-detected from ~/.aws/config or defaults to us-east-1)
  • Bedrock API key (saved to ~/.nemoclaw/credentials.json)

The setup then fetches your available models dynamically and lets you pick one.

Non-Interactive Install (CI/CD)

export AWS_BEARER_TOKEN_BEDROCK="ABSK..."
export AWS_REGION=us-east-2
export NEMOCLAW_PROVIDER=bedrock
export NEMOCLAW_MODEL=us.anthropic.claude-3-5-sonnet-20241022-v2:0
./govclaw-install.sh --non-interactive

Connect and Chat

# Connect to the sandbox
govclaw my-sandbox connect

# Inside the sandbox, open the TUI
openclaw tui

# Or send a single message via CLI
openclaw agent --agent main --local -m "hello" --session-id test

Inference Providers

Provider Auth Notes
AWS Bedrock Bedrock API key (AWS_BEARER_TOKEN_BEDROCK) Native bedrock-converse-stream. Dynamic model discovery.
NVIDIA Endpoints NVIDIA_API_KEY Curated models on integrate.api.nvidia.com.
OpenAI OPENAI_API_KEY GPT models.
Anthropic ANTHROPIC_API_KEY Claude models.
Google Gemini GEMINI_API_KEY Gemini models via OpenAI-compatible endpoint.
Local Ollama None Local models on localhost:11434.

Bedrock Model Access

Bedrock only shows models you have access to. To enable more models:

  1. Go to Bedrock Model Access in your region
  2. Click Manage model access
  3. Enable the models you want (Claude, Llama, Nova, etc.)

Policy Presets

GovClaw includes policy presets for common integrations. Apply during setup or later with govclaw <name> policy-add.

Preset Description
pypi Python Package Index access (suggested)
npm npm and Yarn registry access (suggested)
mattermost Mattermost API and webhook access
discord Discord API, gateway, and CDN
slack Slack API and webhooks
telegram Telegram Bot API
docker Docker Hub and NVIDIA container registry
huggingface Hugging Face Hub, LFS, and Inference API
jira Jira and Atlassian Cloud
outlook Microsoft Outlook and Graph API

Key Commands

Command Description
govclaw onboard Interactive setup wizard
govclaw list List all sandboxes
govclaw <name> connect Shell into a running sandbox
govclaw <name> status Sandbox health and inference status
govclaw <name> logs --follow Stream sandbox logs
govclaw <name> destroy Delete sandbox
govclaw <name> policy-add Add a policy preset
govclaw <name> policy-list List presets (applied/available)

The nemoclaw command also works as an alias.


Architecture

Host
  └── Docker
       └── OpenShell Gateway
            └── k3s
                 └── GovClaw Sandbox Pod
                      └── OpenClaw Agent + GovClaw Plugin
                           └── bedrock-converse-stream → Bedrock Runtime

For Bedrock, OpenClaw uses its native bedrock-converse-stream API to talk directly to AWS Bedrock from inside the sandbox. The network policy allows *.amazonaws.com with TLS passthrough.


Upstream Compatibility

GovClaw is designed to stay mergeable with upstream NemoClaw:

  • Bedrock is integrated as a REMOTE_PROVIDER_CONFIG entry, following the same pattern as OpenAI, Anthropic, and Gemini
  • GovClaw-specific files (govclaw-install.sh, bin/govclaw.js) don't modify upstream files
  • Feature additions (Mattermost preset, policy fix) are additive and contribution-ready

See CHANGES.md for a full changelog.


Learn More

License

This project is licensed under the Apache License 2.0.

About

Claw built for Government. Run OpenClaw inside NVIDIA OpenShell with managed inference and security monitoring

Resources

License

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages

  • JavaScript 45.3%
  • Shell 32.2%
  • TypeScript 17.1%
  • Python 4.6%
  • Other 0.8%