Skip to main content

AI Gateway: One Pi Controls Every PLC

James M. Belcher Founder, JMB Technical Services LLC April 2026 | ControlForge v1.0.541

1. Overview​

A single ControlForge instance — running on a $35 Raspberry Pi — can serve as an AI gateway to every ControlForge on your network. Your AI assistant (Claude Code, Cursor, Windsurf, or any MCP-compatible tool) connects to that one MCP server, and from there it can reach any ControlForge instance just by changing the host and port on each tool call.

No cloud. No relay. No separate server install. The same binary that runs your PLC programs also serves the MCP tools.

Why This Matters​

  • One connection, entire plant. Your AI assistant opens one MCP session and can read variables, deploy programs, debug issues, and build HMI dashboards on any ControlForge instance on the network.
  • Zero infrastructure. No MQTT broker, no cloud account, no VPN, no Docker. Just a ControlForge binary on a Pi.
  • Cross-platform. The gateway Pi can manage Linux, Windows, and ctrlX instances — ARM and x86 — all from the same MCP session.
  • 80 tools per target. Programs, tasks, variables, HMI, debugging, protocol analysis, fleet discovery, and more. All accessible on every target.

2. Hardware You Need​

ItemPurposeCost
Raspberry Pi (3B+ or newer)Runs the MCP gateway~$35
SD card (8GB+)Pi OS + ControlForge binary~$8
Ethernet or Wi-FiNetwork access to other ControlForge instances—

The gateway Pi doesn't need to run PLC programs itself (though it can). Its primary job is translating AI tool calls into HTTP API requests to your other instances.

Any Linux machine on the network works — the Pi is just the cheapest dedicated option.

3. Setup (5 Minutes)​

3.1 Install ControlForge on the Pi​

Download the ARM64 Linux tarball and extract it:

# On the Pi
mkdir ~/controlforge && cd ~/controlforge
tar xzf goplc-linux-arm64.tar.gz

The tarball is self-contained — binary, Node-RED, web IDE, everything.

3.2 (Optional) Start a ControlForge Runtime on the Pi​

If you want the Pi itself to also run PLC programs:

./start-controlforge.sh
# Starts on port 8082, installs systemd service for auto-start on reboot

If the Pi is gateway-only, skip this step. The MCP server doesn't need a local runtime — it talks to remote instances over HTTP.

3.3 Add the MCP Server to Your AI Tool​

On your workstation (where you run Claude Code, Cursor, etc.):

Claude Code:

claude mcp add controlforge -- ssh pi@10.0.0.170 /home/pi/controlforge/controlforge mcp

This tells Claude Code to SSH into the Pi and launch the MCP server over stdio. Every tool call flows through that SSH tunnel.

If Claude Code runs on the Pi itself:

claude mcp add controlforge -- /home/pi/controlforge/controlforge mcp

Other MCP clients (Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "controlforge": {
      "command": "ssh",
      "args": ["pi@10.0.0.170", "/home/pi/controlforge/controlforge", "mcp"]
    }
  }
}

3.4 Verify It Works​

Restart your AI tool, then ask it:

"Get the info from my ControlForge on 10.0.0.31 port 8082"

The AI will call controlforge_info(host="31", port=8082) and return version, uptime, license status, program count, and more.

4. Talking to Multiple Instances​

Every MCP tool accepts host and port parameters. To talk to a different ControlForge, just change them:

# Read variables on the Pi
controlforge_variable_list(host="170", port=8082)

# Deploy a program to the Windows PC
controlforge_deploy(host="31", port=8082, task="MainTask", programs=[...])

# Check runtime status on a ctrlX controller
controlforge_runtime_status(host="45", port=8082)

# Debug a program on a second Pi
controlforge_debug_enable(host="171", port=8082)

Host Shorthand​

The host parameter supports shorthand — just use the last octet:

You typeExpands to
"170"10.0.0.170
"31"10.0.0.31
"45"10.0.0.45
"10.0.1.50"10.0.1.50 (used as-is)
"localhost"localhost

Discover All Instances​

Use fleet discovery to find every ControlForge on your network:

controlforge_fleet_discover(host="170", port=8082)

This returns all reachable instances with their IP, port, version, and status — giving your AI a map of the entire plant.

5. What Your AI Can Do Across the Fleet​

With the gateway in place, your AI assistant has full access to every ControlForge. Here are real workflows that work across instances:

5.1 Deploy the Same Program Everywhere​

"Deploy this temperature controller to all three PLCs"

The AI calls controlforge_deploy three times with different hosts — same program source, different targets. One conversation, entire fleet updated.

5.2 Monitor Variables Across Machines​

"Show me the pressure readings from all my PLCs"

The AI calls controlforge_variable_get on each instance and presents a unified view.

5.3 Cross-Instance Debugging​

"The output on 10.0.0.45 seems wrong — check what its inputs look like vs 10.0.0.170"

The AI reads variables from both instances, compares them, and identifies the discrepancy.

5.4 Fleet-Wide Configuration​

"Set the scan time to 50ms on all instances"

The AI calls controlforge_task_update on each target with the new scan time.

5.5 Build HMI Dashboards Per Machine​

"Create a pump status dashboard on the Pi and a motor dashboard on the Windows PC"

The AI calls controlforge_hmi_create on each target with machine-specific HTML.

6. Authentication​

If any ControlForge instance has JWT authentication enabled, set the token as an environment variable before the MCP server starts:

export CONTROLFORGE_AUTH_TOKEN="your-jwt-token-here"

The MCP server includes this as a Bearer header on every HTTP request. All instances that share the same auth token work automatically.

For instances with different tokens, configure them in your ControlForge config file (see the Configuration Guide).

7. Network Considerations​

Firewall​

Each ControlForge instance must have its API port open. On Linux:

sudo ufw allow 8082/tcp

Latency​

The MCP server makes HTTP calls to each target. On a local network, round-trip times are typically 1-5ms. Even over a slower link, tool calls complete in under a second.

No Internet Required​

Everything runs on your local network. The AI assistant connects to the MCP server, the MCP server connects to your PLCs. No data leaves your network.

8. Example: Complete Home Automation Setup​

Here's a realistic setup using the gateway pattern:

DeviceLocationRoleAddress
Pi 3B+Network closetAI Gateway + HVAC controller10.0.0.170:8082
Pi Zero 2WGarageGarage door + lighting10.0.0.171:8082
Pi 4WorkshopCNC + dust collection10.0.0.172:8082
Old laptopOfficeMonitoring dashboard10.0.0.31:8082

One claude mcp add command. The AI can now:

  • Write HVAC logic for the closet Pi
  • Deploy garage door safety interlocks to the Zero
  • Monitor CNC spindle RPM from the workshop Pi
  • Build a master dashboard on the laptop showing all four systems

All from natural language conversations.

9. Troubleshooting​

"Connection refused" on a target​

The ControlForge instance isn't running or the port isn't open:

# Check if ControlForge is running on the target
ssh pi@10.0.0.170 "ss -tlnp | grep 8082"

# Open the firewall if needed
ssh pi@10.0.0.170 "sudo ufw allow 8082/tcp"

MCP tools don't appear in Claude Code​

Restart Claude Code after adding the MCP server:

claude mcp add controlforge -- ssh pi@10.0.0.170 /home/pi/controlforge/controlforge mcp
# Then restart Claude Code

SSH connection drops​

Use an SSH key for passwordless auth (the MCP transport needs non-interactive SSH):

ssh-copy-id pi@10.0.0.170

Wrong version on a target​

Check from the AI:

controlforge_info(host="170", port=8082)

If the version is old, update the binary on that machine and restart.

10. Summary​

WhatHow
InstallExtract tarball on a Pi
Connect AIclaude mcp add controlforge -- ssh pi@IP /path/to/controlforge mcp
Talk to any PLCChange host and port on any tool call
Discover fleetcontrolforge_fleet_discover
Cost$35 Pi + ControlForge binary
DependenciesNone. Single binary. No cloud.

One Pi. One binary. One MCP session. Every PLC on your network, controlled by AI.