Hopper: The First AI Agent That Drives TN3270 and z/OS Itself
The mainframe is the ghost in the machine of modern finance. It quietly clears payments, books flights, adjudicates insurance claims, and stores decades of business logic that nobody on staff is fully qualified to touch anymore. And while every other corner of software has been swallowed by AI coding agents, the mainframe has remained almost untouched — a green-screen world of TN3270 terminals, ISPF panels, column-72 JCL, and JES return codes that AI agents fundamentally do not know how to operate.
Hypercubic, a startup founded by alumni of Cognition, Windsurf, Apple, Deloitte, and Volkswagen, just changed that. On May 12, 2026 the company released Hopper, which it bills as the first agentic development environment for the mainframe. It's free for individual developers, ships for macOS, Windows, and Linux, and as of this writing is on version 1.1.6.
"Hopper is not trying to hide the mainframe behind a generic abstraction. It is not a chatbot pasted onto a terminal. It is built around the objects and workflows mainframe teams use every day." — Hypercubic, launch announcement
That last sentence is the entire pitch, and the entire reason this matters. Let's unpack what's actually inside.
Why the mainframe resisted AI agents for so long
If you've never touched a mainframe, the easiest way to picture the problem is this: imagine if every codebase still ran inside a fixed 80-column terminal, files were stored in something called a partitioned dataset instead of a folder, every build script was column-sensitive JCL where a stray character past column 71 silently broke the job, the build output came back as a numbered spool file routed by a JES2 queue, and the most important error messages were abend codes like S0C7 or U4038 that mean nothing without context.
Now imagine that this environment is what runs the back office of a major U.S. bank, and the people who know it best are retiring faster than they can be replaced.
That's the operational reality at most Fortune 500 financial, government, and insurance shops. Mainframes still process the bulk of the world's high-volume transactional workloads — and the tooling looks like 1985 because the contracts look like 1985: 24/7 uptime, zero-defect releases, signed-off change windows.
Modern AI coding agents like Cursor, Claude Code, and Codex were built on top of three assumptions: a Git-style file tree, a Unix-like shell, and an HTTPS API surface. A z/OS LPAR offers none of those. Datasets aren't files. JCL isn't a shell script. CICS isn't an HTTP server. Hopper's job is to translate that whole stack into something an agent can drive.
What's actually in the box
Hopper looks, at first glance, like a slightly opinionated TN3270 emulator with a chat panel attached. That is misleading. The chat panel is the lead instrument. The terminal is the fallback.
The agent ships with over 50 mainframe-specific tools, per Hypercubic's launch post. They cover the surfaces that traditionally require an experienced operator:
| Surface | What the agent can do |
|---|---|
| TN3270 / ISPF | Navigate panels by ID, read screens, press PF/PA keys |
| Datasets / PDS | List, filter, read members, edit in place |
| JCL | Write column-strict job control, submit, resubmit |
| JES spool | Parse JESMSGLG, JESYSMSG, SYSUDUMP into structured diagnostics |
| VSAM | Query records as if they were SQL tables |
| CICS | Run transactions, perform NEWCOPY rollouts |
| Return codes | Decode abends into failing step + source line |
The differentiator versus "a chatbot next to PuTTY" is that the agent probes the LPAR on connect. According to Hypercubic, the first time you point Hopper at a mainframe, it discovers your dataset high-level qualifiers (HLQs), your JES configuration, your compile procedures, and your CICS region. Every subsequent action it takes is shaped by your shop's specific conventions, which is critical because mainframe environments are notoriously local: two banks can run the same software with subtly incompatible JCL idioms.
There's also a small but important UX detail: @-tagging. From the datasets panel or jobs panel, Send to chat attaches a dataset, PDS member, or JOBID to the prompt as a structured reference. You don't paste source code into the agent. You hand it a pointer, and the agent reads the resource itself with full fidelity, including EBCDIC encoding and fixed-width formatting. This is the same pattern Cursor uses with @filename, but adapted to mainframe primitives.
A worked example: triage a failed batch job
Hypercubic's launch demo runs through a workflow that every mainframe shop will recognize. A batch job fails overnight. The on-call developer used to:
- Open SDSF, find the failed JOBID.
- Scroll through
JESMSGLGlooking for the abend code. - Cross-reference the code in IBM's message manuals.
- Walk back to
JESYSMSGto find which job step failed. - Open the JCL member, locate the relevant
EXEC PGM=line. - Open the source PDS, find the offending line.
- Patch, resubmit, wait.
That sequence is hours of work in production environments, and the institutional knowledge required to do it quickly is exactly the thing leaving the building as senior operators retire. In Hopper, the developer tags the JOBID into chat and types "Why did this fail?" The agent reads the spool, identifies the abend, traces it to the failing step, opens the source line, proposes a JCL fix, and — pending approval — resubmits.
The pending approval part is important. Sensitive actions require explicit user confirmation. The TN3270 terminal stays visible at all times. The agent waits when you grab the keyboard. This is the design choice that makes Hopper viable inside change-control regimes that would never accept a fully autonomous agent loose on a production LPAR.
How Hopper got here: from HyperFrame to TN3270
Hopper isn't the first thing Hypercubic built. Per the company's own origin story, the project started as a CLI called HyperFrame that connected to a mainframe over FTP, shuttled code and jobs between a local workstation and the LPAR, and made it possible to run COBOL programs through JCL from a normal developer workflow.
That was useful but limited. FTP only sees datasets — it doesn't see screens. Anything that lives behind ISPF panels (and on a mainframe, that is most things) was invisible to the agent. So Hypercubic added TN3270 protocol support, which is the moment the architecture flipped: instead of treating the mainframe as a remote filesystem, Hopper began driving the same green-screen interface humans use. That's what made it agentic rather than just remote-control.
The naming is also deliberate. Hopper is named after Rear Admiral Grace Hopper, who developed the first compiler (the A-0 system) in 1952 and championed the work that produced COBOL. Hypercubic's framing is that Grace Hopper made computing accessible to humans, and Hopper-the-IDE extends that mission to AI agents.
Pricing, packaging, and the enterprise question
There are two tiers:
- Hobby: Free. No credit card. macOS (Apple Silicon and Intel), Windows (x64), and Linux (Debian/Ubuntu .deb, AppImage). All core features. You bring your own mainframe — or request free shared-mainframe credentials from Hypercubic, powered by Maintec.
- Enterprise: Custom pricing. Adds SAML SSO, MCP Server access, admin and model controls, organization-wide privacy controls, a no-training guarantee on customer data, priority support, on-prem and VPC deployment, and SOC 2 + pen-test reports.
The Enterprise list is the giveaway that Hypercubic is going after regulated industries first. Mainframe shops aren't going to install anything that phones home with COBOL source over the public internet. The on-prem and MCP-server posture is what makes Hopper sellable into banks, payment networks, and government systems.
The MCP angle is also worth flagging. Hypercubic's product copy mentions Model Context Protocol prominently, which means Hopper isn't just shipping its own agent — it's exposing the mainframe as an MCP server that other agents (Claude Code, Cursor, custom internal agents) can connect to. That positions Hopper as infrastructure, not just a tool.
What this doesn't fix
A few honest caveats, because the hype here will get loud.
First, Hopper does not modernize anything by itself. It makes the mainframe operable by agents. The actual lift-and-shift work — converting COBOL to Java, replatforming CICS transactions, untangling VSAM — sits in Hypercubic's other products (HyperDocs for system explanation, HyperTwin for tribal knowledge capture, HyperLoop for transformation). Hopper is the operating layer for that platform, not a replacement for it.
Second, the agent is only as good as the LPAR's conventions are legible. Shops with disciplined dataset naming and clean JCL will get fast wins. Shops with 40 years of accreted, undocumented conventions — and that's most of them — will need to invest in cleanup before Hopper's autonomy is comfortable.
Third, autonomy ceilings remain real. Mainframe outages are career-ending. No serious shop is going to let an agent push to a CICS production region without human approval, no matter how good the demo looks. Hopper acknowledges this with mandatory approvals on sensitive operations, but it means the productivity ceiling for the next 12–24 months is augmentation, not replacement.
The competitive picture
Hopper isn't entering empty space. Microsoft has been publishing Azure-based AI agent pipelines for COBOL-to-Java migration, including an open-source samples repo (Azure-Samples/Legacy-Modernization-Agents) that wires up the Microsoft Agent Framework for analysis, conversion, and dependency mapping. IBM has its own mainframe modernization story tied to watsonx Code Assistant. And consultancies — Deloitte, Accenture, Kyndryl — have built large practices around exactly this problem.
What Hopper does differently is operate inside the live mainframe terminal instead of treating modernization as an offline batch job. The Microsoft samples convert code from extracts. IBM's tooling focuses on code transformation. Hopper sits at the operator's seat. That's a smaller surface — but it's the surface where developer flow actually happens.
The Bottom Line
Hopper is the first credible attempt to put a coding agent inside the green screen rather than next to it. It won't single-handedly retire the mainframe, but it lowers the daily cost of working with one — and that's a much bigger deal than it sounds. The mainframe modernization market has been stuck in consulting-engagement mode for two decades precisely because the day-to-day operator loop is so painful that nothing else can get traction. Hopper attacks that loop directly.
The version 1.1.6 release is early. The free Hobby tier is generous enough to actually try. And if you maintain anything that runs on z/OS, this is the first AI tool that meets you where you actually work — column 72, JES queue, abend code, and all.
