Spawning The Claw Deck: Running Autonomous AI Agents on Fedora Linux

One of the primary reasons I walked away from Windows was developer friction. Running background AI agent daemons, custom CLI bridges, and long-running Docker containers on Windows was an endless battle against sleep timers, socket mapping bugs, and memory bloat.

On Fedora Linux, the goal was simple: turn this Razer Blade 14 into an autonomous mission control where multiple AI agents can live, execute code in secure sandboxes, and collaborate in real-time.

Welcome to The Claw Deck.

1. The Architecture: NemoClaw, OpenShell & Hermes

Instead of letting an agent run wild with raw, unrestricted terminal access on the host system, we use NemoClaw and NVIDIA OpenShell.

This architecture isolates each agent in its own lightweight container with granular egress firewall policies: - Sandbox Isolation: The agent gets its own virtual filesystem, package managers, and environment. - Network Presets: Egress traffic is strictly controlled through policy presets (e.g. github, npm, pypi, public-reference, and discord), preventing unauthorized network calls or accidental data exfiltration. - Inference Engine: We chose MiniMax M3 hosted on NVIDIA NIM for Hortense. MiniMax M3 gives us long-context retention and strong tool-calling reliability inside the Hermes agent runtime.

2. Real-World Onboarding Hurdles (And How We Fixed Them)

Setting up bleeding-edge multi-agent infrastructure on a fresh Linux install rarely happens in a single click. Here were the two critical hurdles we hit and resolved:

Hurdle 1: The Non-Root Docker Socket Permission Trap When NemoClaw started provisioning the container sandbox for Hortense, the onboarding wizard abruptly crashed with an unhelpful gRPC HTTP/2 stream reset:

tonic::transport::Error(Transport, hyper::Error(Http2, Reset(StreamId(3), PROTOCOL_ERROR)))

Digging into the daemon logs revealed the root cause: Docker was running, but our user account hadn't been added to the docker group. OpenShell was getting an immediate permission denied /var/run/docker.sock and dropping the gRPC pipe.

The Fix:

sudo usermod -aG docker $USER
sudo systemctl restart docker
newgrp docker

Hurdle 2: The Gateway Version Drift Conflict After resolving Docker permissions, running nemoclaw onboard --resume hit a version mismatch error:

Refusing the system OpenShell gateway service: /usr/bin/openshell-gateway is 0.0.111, 
outside the maximum 0.0.101 supported by this NemoClaw release.

A system-level RPM package had installed a newer build of openshell-gateway (0.0.111) in /usr/bin/, while NemoClaw expected its pinned 0.0.101 binary in ~/.local/bin/.

The Fix: Removing the system-level RPM allowed NemoClaw to use its bundled, compatible binaries cleanly:

sudo dnf remove openshell openshell-gateway -y
systemctl --user restart nemoclaw-openshell-gateway.service
nemoclaw onboard --resume

3. The Claw Deck: Centralized Discord Mission Control

With the sandbox and inference route active, we connected Hortense to Discord.

Instead of managing separate CLI windows for every agent, we created a private Discord server called The Claw Deck.

Why Discord is the Ideal Agent HQ: - Dedicated Channels per Agent: #hortense for general coding and analysis, #gwen for research navigation, and #war-room for multi-agent collaboration. - Syntax Highlighting & Threads: Native code blocks in Go, Python, and TypeScript, with isolated task threads that keep channels clean. - Mobile Access: Ability to assign tasks, check status, or trigger workflows from a phone on the go. - Team Identity: We even designed a custom server emblem—5 cybernetic mechanical lobster claws meeting in a team huddle with glowing NVIDIA green circuitry.

4. What's Next for the Fleet

Hortense is officially online, listening on Discord, and executing tasks inside her OpenShell sandbox.

Next up on The Claw Deck: 1. Spawning Gwen, our dedicated web navigation and research agent. 2. Building an automated CodeAudit reviewer that hooks directly into GitHub pull requests. 3. Enabling local GPU inference using our laptop's RTX 4070 for zero-latency, offline tasks.

Linux has transformed this machine from a simple laptop into an autonomous AI workshop.