Inspiration
I worked at a law firm that still ran its software through a terminal emulator. Screens like that were designed about fifty years ago, and they still run billing, payroll and orders at a lot of organizations. The people who know how to ask those systems a question are retiring, while everyone else now works in tools like Jira. So the answer to an incident often sits on a mainframe that almost nobody on the team can read.
Mainframe Desk connects the present to the past: the team stays in Jira, and the mainframe answers there.
What it does
Mainframe Desk is a Forge app with a Rovo agent. On any Jira or Jira Service Management work item you ask a question in plain English, for example "Why did job BILLNITE fail?". The agent picks one of the lookups the mainframe's administrator has defined, a gateway signs on to the 3270 mainframe, runs that lookup, reads the screen, and the agent answers in seconds. The result is also recorded on the work item as a comment containing the named fields and the mainframe's own screen text, so there is evidence and not just an AI summary. In service-desk spaces it is posted as an internal note.
In the hosted demo the agent can look up batch job history, order status, dataset attributes and the system clock on a real MVS 3.8j system.
Safety is the core of the design:
- The AI never sees a mainframe password. Sign-on scripts and credentials stay on the gateway.
- The agent cannot type free-form keystrokes. It can only run named operations with validated parameters that the administrator wrote.
- Read-only lookups and write operations are separate actions, and the gateway refuses a write on the read-only path.
- Nothing on the mainframe changes: no rewrite, no migration, no new software on the host.
The same operations are exposed as an MCP server from the Forge app, so any other Rovo agent can use them as tools.
How I built it
- Forge app: a
rovo:agentmodule with actions for listing operations, read-only lookups and write operations, arovo:mcpmodule exposing the same tools, and a Jira admin page (UI Kit) for connecting your own gateway. Settings live in Forge storage. - Gateway: a small Node.js service that drives
s3270. Each operation is a scripted screen sequence with typed parameters and regular expressions that pull named fields off the final screen. In demo mode it only accepts calls carrying a Forge Invocation Token for this app. - Demo mainframe: MVS 3.8j (TK4-) on the Hercules emulator, with small TSO programs and data I wrote for the demo, reseeded automatically whenever the system restarts.
Challenges I ran into
- Green screens are stateful. One interrupted session can leave a user signed on and block the next lookup, so sign-off and recovery took more work than the lookups themselves.
- Getting the agent to run a lookup once, not twice, and to record a result exactly once per question.
- Making the judge experience zero-setup: a fresh install talks to a hosted demo mainframe with nothing to configure.
Accomplishments that I'm proud of
A real mainframe answers a question asked in Jira in about half a minute, with the raw screen attached as proof, and the AI never holds a credential.
What I learned
The hard part of legacy integration is not the protocol. It is deciding exactly what an AI is allowed to do on a system of record, and making that list short, named and reviewable.
What's next for Mainframe Desk
More reliable sign-on handling, IBM i (5250) hosts, per-operation permissions by Jira role, and a Marketplace listing.
