The GS UK Virtual Conference, which happened on April 21-23, brought together voices from across the open source and mainframe ecosystem for a day of insight, collaboration and forward-looking discussion. For the third year in a row, Open Mainframe Project had a stream at the event. Sessions highlighted how organizations are modernizing critical infrastructure, advancing open source innovation on the mainframe and building stronger technical communities through shared expertise and real-world experience.
At GSUK 2026, Joe Winchester, Senior Technical Staff Member at IBM and Open Mainframe Project Ambassador, walked through something that felt less like a conference demo and more like a preview of how mainframe work is going to get done. The session: connecting Zowe to an MCP server and letting an AI agent handle the heavy lifting.
The result was a live demonstration of what’s possible when decades of mainframe tooling meets the Model Context Protocol, and a strong case for why this combination is the next major productivity shift for the platform.
What Is MCP, and Why Does It Matter for Mainframe?
MCP stands for Model Context Protocol. Think of it as a universal connector standard for AI: the USB-C of the AI world. It gives AI assistants a structured way to call external tools, access data sources, and take actions. Instead of a large language model that can only respond to text, you get an agent that can actually do things.
For mainframe, that means an AI that lists datasets, submits jobs, and takes action, using tools built on top of Zowe.
Zowe CLI as the Foundation
If you know your way around a terminal, Zowe CLI lets you interact with z/OS from any machine (Mac, Windows, Linux) without a 3270 session. You can list datasets, view members, submit jobs, issue TSO and console commands. It’s powerful, but it requires precision. You have to know the commands, the syntax, the dataset names. That’s the gap MCP agents close.
Guardrails Are Part of the Design
One thing Joe emphasized throughout: the MCP server isn’t a raw connection to z/OS. Tools are classified by risk. Read-only operations run without prompting. Operations with potential side effects (like job submission) require confirmation. More destructive operations have additional checks.
The goal is an agent that’s genuinely useful for day-to-day work without becoming a runaway process. The community is actively working out where those guardrails need to be, and input from experienced system programmers is specifically what the project needs right now.
Use Cases Worth Watching
Joe walked through several categories where this combination shows clear value:
Developer onboarding. New mainframers can ask what a dataset is, what a job does, what a command will produce, and get answers grounded in the actual system they’re connected to, not just documentation.
Overnight job failure triage. An agent that can read syslog output, identify abends, explain what went wrong, and suggest or execute fixes is a meaningful shift for operations teams.
Legacy code understanding. The ability to read a JCL or COBOL program, explain what it does, and propose modernization paths is an emerging capability the community is exploring.
Compliance and audit support. System state queries, job history, and configuration checks become conversational rather than requiring specialist CLI knowledge.
Get Involved
The Zowe MCP server is open source and at an early stage where community input shapes its direction. The repository is at github.com/zowe/zowe-mcp. Raise issues, try the mockup backend, or connect on the Open Mainframe Project Slack.
If you’re building extensions for Zowe CLI — Broadcom, IBM, Rocket, and others have contributed plugins over the years — the extensibility model for MCP tools is being designed with that ecosystem in mind.
Keep up to date with Open Mainframe Project:
- Connect with our LinkedIn page
- Sign up to get our quarterly newsletter
- Subscribe to our Youtube Channel
- Bookmark our Flickr channel
- Follow us on X at @Openmfproject
