Universal OData ↔ MCP Bridge; Or: How I Accidentally Made 15,000 Enterprise Services Talk to AI

[Alice_V]

Or: How I Accidentally Made 15,000 Enterprise Services Talk to AI

In March 2025 I realised the "hard part" of enterprise AI integration was already solved by metadata we've had for years. Two hundred MCP servers suddenly became 15,000+ – with one simple translator. 🚀

What Problem Are We Solving?

Hey folks! 👋 This is the logical continuation of my ABAP-native AI journey:

• ZVDB & ZLLM brought AI into ABAP (vector search, LangChain-lite, local inference)

• This bridge brings ABAP — and every other OData system — to AI

Enterprises sit on thousands of OData endpoints that already expose clean business semantics, yet modern AI assistants can't reach them. All that was missing was a translator.

The "15,000+ MCP Servers" Moment

Ecosystem	OData services made AI-ready
SAP (ECC → S/4 & BTP)	≈ 10,000
Microsoft Dynamics	≈ 1,500
Salesforce	≈ 800
Oracle Cloud ERP	≈ 2,000
SharePoint, ServiceNow & friends	≈ 1,000

Total: 15,000+ enterprise-grade services now surface as first-class MCP tools.¹ Not toys—systems that move real money.

TL;DR (Four Bullets, Promise!)

1. Code → GitHub: oisee/odata_mcp (Python) & oisee/odata_mcp_go (Go) (link in first comment)

2. Problem → $2T of software speaks OData; 0% speaks MCP

3. Solution → Bridge auto-discovers every entity/action; zero config

4. Result → Your ERP landscape is AI-accessible in under two minutes

Why This Matters—The "So What?"

Role
CTO / Architect	All legacy & cloud systems become AI-ready today—no vendor roadmap, no middleware
Developer	One interface for SAP, Dynamics, Oracle, Salesforce—write once, chat everywhere
Business user	"Claude, approve all POs < $50k" – and it just happens

The "Wait, That's It?" Moment

• OData = REST + machine-readable metadata

• MCP = JSON-RPC + tool definitions AI assistants understand

• Bridge = translate the metadata to tool definitions. That's it.

➡️Just translate between them—and you've unlocked the whole enterprise. 🤔

Two Flavours, Same Power

	Python	Go
Startup	~3s	~0.3s
Memory	~150 MB	~30 MB
Best for	Prototyping & demos	Production workloads
Shared superpowers	Full CRUD, function imports, CSRF juggling (hi SAP 👋), Basic auth, Cookies auth

Demo: From Skepticism to "Holy Wow"

Enterprise Reality Demo

The Auto-Discovery Magic ⭐

OData's metadata does 90% of the work; the bridge just speaks fluent MCP.

Want the deep dive?

Check the repo or catch me on SAP On Azure Podcast 🎙️

The Hidden Cost of Waiting

	Vendor AI (Joule / Copilot / Einstein)	Bridge Today
Availability	Cloud-only, phased rollout	Any OData v2/v3—even ECC 2007
Coverage	Single vendor	Entire landscape
Time-to-value	6-18 months	< 2 minutes
Data location	Vendor cloud	Stays on-prem / your VPC
Vendor-Lock	High	Zero (MIT licensed)

Security & Governance (Because Enterprise)

• Runs inside your network — nothing leaves unless you proxy it

• Honours existing roles — if users can't see it in GUI or Fiori - they can't see it via MCP

• Audit trail-ready — every request loggable with full context

No new attack surface—just a protocol translator.

The Ecosystem Just Shifted

When 15,000+ production services become AI-accessible overnight, MCP stops being a dev toy and becomes the enterprise standard. This isn't iteration—it's inflection.

A Day in Your New AI-Powered Life

09:00  "Claude, what needs my attention?"
       → 3 POs > $40k awaiting approval (SAP)
       → Customer escalation (ServiceNow)
       → Budget variance on cost center 4210 (Dynamics 365)
09:05  "Approve the POs from established vendors, escalate to Sarah, schedule variance review."
       → Done. Next?

Five minutes. One assistant. Zero UI clicks.

What Could Go Wrong? (Let's Be Honest)

• Query complexity: Some OData implementations choke on deeply nested $expands

• Rate limits: Your backend might not love 100 requests/second from an enthusiastic AI

• Schema changes: When your OData model changes, tools need regeneration

• Governance: "Claude, show me all salaries" works if your auth is broken

But these are solvable with basic ops hygiene—rate limiters, caching, and proper RBAC.

Get Rolling in Two Minutes

1. download pre-built binary for your platform: Release Universal OData ↔ MCP Bridge v1.0.0 (Multi-Platform Binaries) · oisee/odata_mcp_go
2. copy/rename to the appropriate folder (like C:/bin/ )
3. update your claude_desktop_config.json correspondingly:

{
    "mcpServers": {
        "northwind-go": {
            "args": [
                "--service",
                "https://services.odata.org/V2/Northwind/Northwind.svc/",
                "--tool-shrink"
            ],
            "command": "C:/bin/odata-mcp.exe"
        }
    }
}

Add as many services as you want: just repeat with different --service argument

Instructions on how to build from source code - in README here: oisee/odata_mcp_go

⭐Star the repo if this unlocks value for you!

What's Next

• Stay tuned for next week's deep dive: "10 Mind-Blowing Things You Can Build When Your ERP Speaks AI"

But honestly? Even this changes everything.

Try It. Break It. Make It Better.

Here's what I need from you amazing people:

1. Try it with your systems - SAP, Dynamics, whatever you've got

2. Break it creatively - I love good bug reports

3. Add your system's quirks - Every OData implementation is a special snowflake

4. Share what you build - The crazier the use case, the better

Drop a PR, open an issue, or just share what happens when you connect your enterprise to AI.

Want This as an Azure Service?

Look, this open-source bridge proves the concept works. But I know what you're thinking - "Great Alice, but what about enterprise deployment, monitoring, compliance, support?"

Fair point. What if this was available as a managed Azure service? Zero setup, enterprise SLA, built-in monitoring, automatic scaling...

Drop a comment if you want that. 😉

P.S. Yes, it handles all the authentication properly. Yes, it respects permissions. Yes, it works with your ancient SAP system from 2007. No, I don't know why this wasn't built years ago either. (Well, actually, I might know.)

P.P.S. First prompt I'll run in production: "Claude, find every place we're leaving money on the table." 💰

¹ Counts derived from SAP annual report (2024), Microsoft Dynamics docs, Salesforce API catalog, Oracle ERP Cloud guides, and the public awesome-mcp-servers list; full methodology in /analysis/service-counts.csv.

#EnterpriseAI #OData #MCP #OpenSource #AIIntegrationCP #OpenSource #AIIntegration