AI Agents on Metalshub: What MCP Actually Connects

A structured commercial context layer serves as the core foundation for enterprise AI in raw materials trading. We recently published our Metalshub AI vision, establishing that model performance matters far less than giving AI systems access to verified, structured business context. Having a context layer only generates real value once an enterprise’s internal tools can securely reach it.

I sat down with our Head of AI Platform at Metalshub, Julien Dedek, to discuss the technical mechanism that makes this connection possible: the Model Context Protocol (MCP).

The Model Context Protocol (MCP) is an open standard designed to connect AI models directly to external data sources. It is not a vendor-locked or proprietary integration.

Thus far, connecting corporate AI to external business software, platforms, and tools has required custom integrations across ERPs, CRMs, and document archives. Each integration had to be built individually per customer, making implementation relatively slow and expensive to repeat. MCP replaces this approach with a single governed interface exposed once. Any MCP client running on customer infrastructure, such as Corporate ChatGPT, Claude, or Microsoft Copilot, can query platform data using standardised calls.

Architecturally, an MCP client, whether deployed on customer infrastructure or used through interfaces like ChatGPT or Claude, communicates directly with an MCP Server operating on Metalshub infrastructure. The server enforces multi-tenant security policies before exposing backend platform data, including live enquiries, negotiation terms, counterparty relationships, and executed contracts.

This setup changes how models retrieve information. Rather than feeding entire documents into a prompt window, the assistant retrieves specific data points on demand.

By connecting your AI tools directly to live Metalshub data, you unlock two powerful capabilities.

Simply chat with your AI to get precise, context-aware insights, such as:

• Spend and Performance Analysis: Ask your AI which raw material has the lowest average offer rate per enquiry, or which supplier is typically the slowest to respond. The AI instantly and securely scans live transaction records to present a structured summary of cost comparisons or response times.
• Counterparty Shortlisting: Let your AI recommend the best suppliers for a new enquiry based on their verified delivery history rather than static lists.

Instruct your AI to handle operational workflows, such as:

• Negotiation Support: Ask your AI to draft a counter-offer based on your specific historical contract terms with a counterparty.

Strict data governance dictates how information flows through MCP. A customer’s commercial history remains isolated within their account boundary. Metalshub does not use your commercial history to train any public language models. Market-level insights or price benchmarks rely exclusively on aggregated and anonymised network data sets.

Enterprise IT administrators maintain complete control over the connection configuration. They choose which internal systems connect, which language models process the data, and whether the connection permits read-only queries or action-based steps.

“Your commercial history is yours,” says Dedek. “It’s never used to train a shared model, and it never leaves the fence.” Access operates on explicit corporate authentication paired with granular permissions. Administrators can modify or revoke data access scopes at any time without impacting core platform operations.

Establishing an MCP connection follows a structured sequence:

1. Deploy an MCP Client: Confirm your enterprise uses an MCP client, such as Corporate ChatGPT, Claude, or Microsoft Copilot, hosted on internal infrastructure.
2. Configure Authorisation: An IT administrator authorises the connection to Metalshub’s MCP server and sets permission scopes, defining read-only or read-write access for specific datasets.
3. Discover Capabilities: The MCP client automatically discovers available endpoints, queries, and structured actions exposed by the MCP server.
4. Run Verification Queries: Execute a test request, such as retrieving negotiation history for a specific material category, to verify data retrieval and security bounds.
5. Validate Action Safeguards: Test action capabilities, such as drafting a counter-offer, confirming that mandatory human approval stops execution before transmission.

The technical roadmap follows a clear progression: Text-2-Platform data capture (live), MCP read access (rolling out), write and action capabilities (in development), and specialised agent workflows like negotiation assistance. Capabilities build systematically through four stages: read, update, act, and write.

Commercial teams preparing for these connections should begin structuring their commercial data today, as historical context cannot be backfilled automatically later. Organisations should select one commercial workflow to digitise end-to-end, select an enterprise-approved language model, and request early access to MCP connections.

A structured context layer generates value when enterprise systems connect to it securely. The Model Context Protocol provides that live connection. Watch our technical setup video or contact the Metalshub team to request early access.

We remain focused on helping industry leaders translate raw commercial activity into measurable bottom-line value. To explore this vision further and see how leading metals and mining enterprises are preparing their data stacks for the AI era, join our upcoming MetalshubTalks episode on “The State of AI for Buying and Selling Raw Materials.” In this session, our leadership team will discuss how to bridge the gap between transactional systems and enterprise AI platforms, demonstrating practical use cases and setting the stage for future AI-driven commercial operations.