MCP — the protocol layer
What is MCP?
MCP stands for Model Context Protocol. It is an open, vendor-neutral protocol published by Anthropic in early 2025. Its goal: let LLMs (Claude, ChatGPT, others) talk to tools such as editors, databases, or browser automations in a uniform, well-specified language — instead of writing a custom integration each time.
ForgeIEC implements the server side of MCP — ForgeIEC Studio is an MCP server. Any MCP-compatible client can connect — the most important options:
- The built-in chat tab in ForgeIEC Studio itself (default, no setup)
- Anthropic Claude Desktop (Mac, Windows, Linux) —
~/Library/Application Support/Claude/claude_desktop_config.jsonor the equivalent under~/.config/Claude/; enter URL + bearer token from ForgeIEC Studio there - VS Code with an MCP extension (e.g. Anthropic’s “MCP”) — register the ForgeIEC MCP server in the extension settings as an external MCP endpoint; then you can call ForgeIEC tools from VS Code chat sessions
- Claude Code (CLI) —
claude mcp add forgeiec https://forgeiec-ws.local:7531 --token <bearer> - ChatGPT (OpenAI) with MCP connector
- Custom clients you write yourself (see next section)
- mcp-inspector (the official debug tool, npm)
In all cases you need two pieces from ForgeIEC Studio: the
server URL (https://<hostname>:7531) and the bearer
token from Preferences → AI → Profile → Bearer Token. With
those, your external client can execute exactly the same MCP
tools as the built-in chat tab.
Who is this section for?
Depending on your role, two entry points:
For application engineers
If you want to call the MCP server of ForgeIEC Studio from your own scripts or tools. Topics:
- How the call sequence works (lifecycle)
- Authentication via bearer token or client certificate
- Reading the tool list
- Examples with
curl - Handling confirmation prompts
- Plug-in API status
For IT + operations
If you deploy, monitor, or integrate ForgeIEC into an existing IT landscape. Topics:
- Network: ports, firewall rules, bind address
- Certificates: server cert, trust store, renewal
- Identities + tokens: bearer per profile
- Log files: location, format, rotation
- Backup: what to save, what is regeneratable
- Monitoring: server_info as a health check for your monitoring
- Troubleshooting: common problems + recipes
Architecture picture
flowchart TB
Client[LLM client
Claude / GPT / Custom]
Studio["ForgeIEC Studio (MCP server)
• Tools (≈80)
• Confirmation State Machine
• Audit log
• Trust store"]
anvild[anvild
PLC runtime]
daemons[bellowsd / tongs-*
HMI + fieldbus]
Client <-->|HTTP+SSE / JSON-RPC| Studio
Studio -->|gRPC| anvild
anvild -->|iceoryx2 SHM| daemonsThe MCP server is embedded in the ForgeIEC Studio process — it shares memory and lifecycle with the running project. Tool calls are therefore very fast (no RPC hop), and the client always sees the current in-memory state, not a stale file mirror.
Spec references
For deeper details:
- MCP protocol (official):
https://modelcontextprotocol.io - MCP inspector (debug tool, npm):
https://github.com/modelcontextprotocol/inspector - ForgeIEC’s own MCP spec: in the source tree at
documentation/architecture/mcp-platform-v1.md— RFC-2119 normative, approx. 2300 LOC, covers all security and federation decisions