Guide
Using TwinEdge with MCP: Server, Client, and Edge
How TwinEdge lets your AI assistants, your existing business systems, and your on-site equipment talk to each other safely.
1. What MCP is
MCP (Model Context Protocol) is an open standard that lets an AI assistant use tools and read data from other software in a safe, controlled way. Think of it as a universal power socket for AI. Any AI assistant that supports MCP (Claude, Microsoft Copilot, ChatGPT Enterprise, and many in-house agents) can plug into any system that offers an MCP connection.
A simple picture
- An MCP server is the socket. It offers tools (things an AI may do) and resources (things an AI may read).
- An MCP client is the plug. It is the AI side that discovers the tools and calls them.
2. The three TwinEdge MCP pieces
TwinEdge uses MCP in three places. Each one solves a different problem for an enterprise that already has data, systems, and equipment.
Socket
TwinEdge MCP Server
Lets your AI assistants use TwinEdge. Ask your company chatbot about a machine and get an answer grounded in the asset model, live readings, and work history.
Plug
TwinEdge MCP Client
Lets TwinEdge's AI helpers use your existing systems: Maximo, SAP, ServiceNow, SharePoint, data lakes, historians. Your data stays where it is.
On site
TwinEdge MCP Edge
A small MCP server running on the TwinEdge OS box next to the equipment. Works with no internet. A technician's laptop assistant can ask the machine directly.
3. TwinEdge MCP Server
What it is: the socket that lets any MCP-capable AI assistant use TwinEdge. It sits in front of the TwinEdge APIs you already run. Nothing new to install on the data side.
Who it is for: plant managers, reliability engineers, facility leads, finance and operations analysts who already use an AI assistant at work and want it to know about the equipment.
What an assistant can do through it
| Tool | What it does | Type |
|---|---|---|
| list_assets, get_asset | "Show me the chillers at the Denver site." | read |
| get_readings | "What was the power draw on Compressor 07 last week?" | read |
| get_asset_health | "Which generators are drifting away from normal?" | read |
| get_open_alarms | "What is alarming right now, and how urgent is it?" | read |
| get_work_history | "When was this pump last serviced, and what was done?" | read |
| search_manuals | "What does the manual say about this fault code?" | read |
| assets_near | "What equipment is within 500 m of this leak?" | read |
| explain_recommendation | "Why did TwinEdge suggest cleaning the condenser?" | read |
| propose_work_order | "Draft a work order to replace the bearing." Goes to the approval queue. | propose |
| propose_inspection | "Schedule a look at Battery Bank A before the test." | propose |
Tool names are illustrative. The exact list is published by the server and can be filtered per customer.
Example conversation
# Plant manager, in Microsoft Copilot Q: Why did Line 3 use 11% more power this month? # Copilot calls TwinEdge MCP Server: get_readings, get_asset_health, explain_recommendation A: Air compressor 07 is using more power to make the same air. Pressure drop across the separator is rising. TwinEdge already suggested inspecting the separator element. Evidence: 14-day trend, physics model, last service 9 months ago. Want me to draft a work order for approval?
4. TwinEdge MCP Client
What it is: the plug on TwinEdge's side. TwinEdge's own AI helpers use it to reach the systems your company already runs, through those systems' MCP servers.
| Your system | What TwinEdge uses it for |
|---|---|
| Maintenance / EAM (IBM Maximo, SAP PM, ServiceNow) | Read asset master data, open work, spare parts, contracts. Send approved work back into the system of record. |
| Historians and data platforms (AVEVA PI, Ignition, Snowflake, Databricks) | Read past readings without copying the whole history into TwinEdge. |
| Documents (SharePoint, Confluence, Google Drive) | Find drawings, manuals, procedures, and past reports for the asset in question. |
| Finance and purchasing (ERP) | Cost of downtime, part prices, energy tariffs, so recommendations include money. |
| Mapping (Esri ArcGIS) | Where assets are, which sites are nearby, which crews cover them. |
| Ticketing and chat (ServiceNow, Jira, Teams, Slack) | Notify the right person and track the follow-up. |
- Data stays home. TwinEdge reads what it needs, when it needs it. No bulk export, no second copy to secure.
- No custom connector per system. If your vendor offers an MCP server, TwinEdge can use it.
- Richer recommendations. A suggestion that knows the part is in stock and costs $340 is more useful than one that does not.
5. TwinEdge MCP Edge
What it is: a small MCP server that runs on the TwinEdge OS box installed next to the equipment. It exposes what the box knows: live tags, the last days of buffered readings, local alarms, and asset setup.
- Works offline. The assistant on a technician's laptop or phone connects to the box over the site network.
- Answers about the machine in front of you. From local storage, in seconds.
- Same rules as the cloud. Read and propose only. The Edge never writes to a controller.
| Tool | What it does | Type |
|---|---|---|
| list_local_assets | "What is this box watching?" | read |
| read_tag, get_recent_readings | "Current oil pressure?" "Last 24 hours of coolant temperature?" | read |
| get_local_alarms | "What tripped overnight?" | read |
| get_link_status | "Is the PLC connection healthy? When did we last sync?" | read |
| propose_note, propose_work_order | "Log what I found; draft the follow-up." Syncs when the link returns. | propose |
| write to controller / change setpoint | Not offered. Ever. | never |
Example: remote generator site, no signal
6. Safety rules (the same everywhere)
AI near equipment must be boring and predictable. These rules apply to Server, Client, and Edge alike.
Two kinds of tools only
Read tools look things up. Propose tools put a draft in the approval queue. There is no third kind. No tool starts a machine, changes a setpoint, or closes a work order on its own.
People approve
Every proposal waits for a person in TwinEdge, exactly as it does today. The AI assistant cannot approve its own suggestion.
One key per company, per assistant
Each connection gets its own scoped key. It can only see that company's assets. Filtering happens in the database, not in the prompt.
Everything is logged
Who asked, which tool ran, what data was returned, what was proposed, and who approved it. Same audit trail as the rest of TwinEdge.
Evidence travels with answers
Every answer carries its sources: readings, model limits, history, documents. If the evidence is thin, the answer says so.
Least data, not most
Tools return summaries and time windows, not whole tables. Assistants get what they need for the question, nothing more.
7. Real-world examples by industry
| Setting | Question asked | MCP pieces | Result |
|---|---|---|---|
| Manufacturing plant | "Which machines on Line 3 are most likely to stop us this week?" | Server + Client (Maximo, ERP) | Ranked list with cause, part availability, and a draft work order for each. |
| Office tower | "Why is the energy bill up, and what would fix it cheapest?" | Server + Client (ERP tariffs, BMS history) | Chiller condenser fouling identified; cleaning proposed with the expected kWh saving. |
| Telecom tower network | "Which sites will run out of backup runtime if the grid fails tonight?" | Server + Edge (per-site boxes) | Sites listed by risk; inspections proposed for the two weakest battery banks. |
| Hospital campus | "Show me the inspection proof for critical power before the audit." | Server + Client (SharePoint) | Evidence package assembled from asset records and documents, ready for review. |
| Data center | "Is the UPS in Hall B really healthy, or just not alarming?" | Server | Health explained against the model and history, not just the alarm list. |
| Field service | "What happened to this pump since I was last here?" | Edge | Local readings, alarms, and notes, offline, in seconds. |
8. How to start
- 1Pick one assistant and one question. For example, Copilot and "why is this asset using more energy?" Connect it to the TwinEdge MCP Server with read tools only.
- 2Add one proposal tool. Turn on propose_work_order. Watch the approval queue for a week. Tune what the assistant is allowed to draft.
- 3Connect one system of record. Point the TwinEdge MCP Client at your maintenance system's MCP server. Recommendations now include work history and parts.
- 4Enable Edge at one site. Give the field team an offline assistant for one location. Collect what they ask; those questions shape the next tools.
- 5Roll out by site, not by feature. Same tools, same rules, more places.
What we ask of the customer
- A named owner for each connected system (who approves the scope of tools).
- An identity provider or key store to issue and revoke per-assistant keys.
- Agreement on the approval path: who can approve which kinds of proposals.
9. Common questions
Does this mean the AI can control our equipment?
No. There are no control tools. Read and propose only. A person approves every action inside TwinEdge.
Do we have to move our data into TwinEdge?
No. The MCP Client reads from your systems in place. Only what TwinEdge needs for an answer is fetched, and it is logged.
Which AI assistants work with it?
Any that support MCP. Today that includes Claude, Microsoft Copilot (via connectors), ChatGPT Enterprise, and most agent frameworks. If your vendor adds MCP support tomorrow, it works with TwinEdge without changes.
What if our maintenance system has no MCP server yet?
TwinEdge keeps its existing direct integrations. MCP is added on top, not instead. As vendors publish MCP servers, you switch over when ready.
Does the Edge need the cloud?
No. It runs on the site box and answers from local storage. Proposals made offline sync when the link returns.
Can we limit what a given assistant sees?
Yes. Each key is scoped to a company and to a list of allowed tools. A finance assistant can be allowed cost data but not live readings; a technician's assistant the other way round.
10. Glossary
- MCP (Model Context Protocol)
- An open standard that lets AI assistants use tools and read data from other software in a controlled way.
- MCP server
- The "socket": a piece of software that offers tools and data to AI assistants.
- MCP client
- The "plug": the AI side that discovers and calls those tools.
- Tool
- One action an assistant may take, such as "look up an asset" or "draft a work order".
- Read tool
- A tool that only looks things up. It changes nothing.
- Propose tool
- A tool that creates a draft for a person to approve. It never carries out the action itself.
- Asset model
- TwinEdge's live picture of a machine: how it should behave, its limits, its history, and its repairs.
- TwinEdge OS
- The software that runs on the small industrial computer installed on site next to the equipment.
- System of record
- The official place where a type of data lives, for example your maintenance system for work orders.
- Scoped key
- A credential that only works for one company and one allowed set of tools.
See MCP on your own equipment.
Bring one asset and the AI assistant your team already uses. We will show you the loop, with the guardrails on.
- Read and propose only
- Human approval
- Scoped keys
- Full audit log