Localhost Bridge

Bridge Docker containers to host localhost services via socat. Solves the #1 networking issue in containerized AI agent deployments: containers can't reach services bound to 127.0.0.1.

localhost-bridge — Connect containers to host localhost services

⚠️ Security & Privileges

This skill requires host-level privileges. It must be reviewed and executed manually by an administrator — never autonomously by an agent.

What it does on the host:

• Creates a systemd service (persistent across reboots) that forwards traffic from a Docker bridge IP to localhost
• Adds a UFW firewall rule scoped to a specific Docker bridge interface
• Requires sudo, Docker daemon access, and socat from your distro's official package repository

Before running any command:

1. Review the generated /etc/systemd/system/socat-<SOURCE_NETWORK>-<TARGET_SERVICE>-<PORT>.service file — confirm ExecStart binds only to the intended Docker bridge IP (172.x.x.1), never 0.0.0.0
2. Review the UFW rule — confirm it targets the correct br-<ID> interface and port
3. After setup, verify the port is NOT reachable from the public network: curl --connect-timeout 2 http://<PUBLIC_IP>:<PORT>/ must fail
4. Test from inside a container before deploying widely

Do not grant an automated agent permissions to run these commands without human approval.

---

The Problem

A service on the host listens on 127.0.0.1 (AI gateway, MCP server, Ollama, database...). A Docker container needs to reach it. localhost inside the container points to the container itself, not the host. Requests either timeout silently (firewall drops packets) or get connection refused.

The Solution

socat listens on the Docker bridge gateway IP and forwards to host loopback. Combined with a scoped firewall rule, this gives containers access without exposing the service externally.

Setup (run manually as admin)

1. Find the Docker bridge gateway IP

bash

# For a specific container
docker inspect <container_name> --format '{{json .NetworkSettings.Networks}}' \
  | python3 -c "
import json,sys
d = json.load(sys.stdin)
for net, info in d.items():
    print(f'{net}: gateway={info[\"Gateway\"]}')"

2. Create a systemd service

Replace <GATEWAY_IP>, <PORT>, <SOURCE_NETWORK>, and <TARGET_SERVICE> with your values.

Naming convention: socat-<source_network>-<target_service>-<port> — source network is the Docker network (consumer), target service is the host service. Self-documenting.

Examples: socat-bridge-gateway-18789, socat-windmill_default-gateway-18789, socat-bridge-ollama-11434

Review the ExecStart line before enabling — confirm it binds to the Docker bridge IP only.

bash

sudo tee /etc/systemd/system/socat-<SOURCE_NETWORK>-<TARGET_SERVICE>-<PORT>.service > /dev/null << 'EOF'
[Unit]
Description=Socat bridge: <SOURCE_NETWORK> -> <TARGET_SERVICE>:<PORT>
After=network.target docker.service

[Service]
Type=simple
ExecStart=/usr/bin/socat TCP-LISTEN:<PORT>,bind=<GATEWAY_IP>,fork,reuseaddr TCP:127.0.0.1:<PORT>
Restart=always
RestartSec=5

[Install]
WantedBy=multi-user.target
EOF

# Review the file before enabling:
cat /etc/systemd/system/socat-<SOURCE_NETWORK>-<TARGET_SERVICE>-<PORT>.service

sudo systemctl daemon-reload
sudo systemctl enable --now socat-<SOURCE_NETWORK>-<TARGET_SERVICE>-<PORT>

3. Add firewall rule (MANDATORY)

Without this, socat listens but packets from the container are silently dropped — causing 30-second timeouts with no error.

Review the bridge ID before applying — a wrong ID can expose services.

bash

# Find the Linux bridge interface for the Docker network
BRIDGE_ID=$(docker network inspect <network_name> --format '{{.Id}}' | cut -c1-12)

# Verify this is the right bridge
ip link show br-${BRIDGE_ID}

# Allow traffic only on that bridge interface
sudo ufw allow in on br-${BRIDGE_ID} to any port <PORT> proto tcp comment "<SOURCE_NETWORK>-<TARGET_SERVICE>-<PORT>"

4. Verify security

bash

# MUST succeed (from inside a container)
docker exec <container_name> curl -s --connect-timeout 5 http://<GATEWAY_IP>:<PORT>/

# MUST fail (from the public network)
curl --connect-timeout 2 http://<PUBLIC_IP>:<PORT>/

Multi-Network Workers

A container can be on multiple Docker networks. Each has its own bridge IP. You need a socat instance + firewall rule for each network the container uses. In practice, one network is usually enough.

Check all networks: docker inspect <container> --format '{{json .NetworkSettings.Networks}}'

Common Use Cases

Host service	Container client	Default port
AI gateway (OpenClaw, LiteLLM)	Workflow orchestrator (Windmill, n8n)	18789
MCP server	Dockerized agent	varies
Ollama	RAG pipeline, agent	11434
PostgreSQL	API server	5432
Redis	Any containerized app	6379

Troubleshooting

Symptom	Cause	Fix
30s timeout, no error	Firewall dropping packets	Add UFW rule on the bridge interface
Connection refused	socat not running	systemctl status socat-<SOURCE_NETWORK>-<TARGET_SERVICE>-<PORT>
Works then stops after Docker restart	Bridge IP changed	Check new gateway IP, update socat bind
socat won't start after reboot	Docker not ready	Ensure After=docker.service in unit file

Alternatives

Depending on your security posture, consider:

• Docker host networking (network_mode: host) — simpler but removes all container network isolation
• Running socat inside a minimal privileged container — avoids host-level systemd changes
• Configuring the host service to bind to the Docker bridge IP directly — no socat needed, but the service must support custom bind addresses
• host.docker.internal (Docker Desktop) — works on Mac/Windows, not reliably on Linux

Prerequisites

Install socat from your distro's official package repository:

bash

sudo apt-get install -y socat  # Debian/Ubuntu
sudo dnf install -y socat      # Fedora/RHEL

References

• Blog post: The Localhost Trap — why this problem exists and why it matters for AI infrastructure
• Source: Casys-AI/casys-pml-cloud
• Docker docs: Packet filtering and firewalls