If you've spent any time on an AI coding tool in the last two years, it was almost certainly built for the languages and runtimes that dominate GitHub: TypeScript, Python, Go, Rust. The mainframe — the platform quietly running the world's banks, airlines, and government benefits systems — has been left almost entirely untouched. Hopper is a direct bet against that neglect: an agentic development environment that drives z/OS the way Cursor drives a Node repo.
I went in skeptical. "AI for mainframes" reads like a slide in a modernization pitch deck nobody asked for. But the product is more concrete than the category suggests, and the problem it targets is real. Here's the honest breakdown for builders deciding whether to spend an afternoon on it.
What Hopper Actually Does
Hopper is not a COBOL chatbot bolted onto a docs site. It's an agent that connects to your mainframe and operates the green screen on your behalf. You describe what you want in plain English; the agent navigates the system, runs commands, submits work, and reads the results back.
The core capabilities, based on the current product:
- ISPF panel navigation by ID. The agent drives ISPF panels directly and issues z/OS commands via natural language — instead of memorizing the panel hierarchy, you ask for the outcome.
- JCL authoring and submission. It writes column-strict JCL (the formatting that breaks every newcomer), submits the job, and parses the JES return codes autonomously rather than making you eyeball SDSF.
- Abend diagnostics. It decodes JESMSGLG, JESYSMSG, and SYSUDUMP output into structured diagnostics. If you've ever scrolled a dump trying to find the actual failure, this is the headline feature.
- VSAM as SQL. Through an MCP-connected interface, you can query VSAM datasets using SQL semantics instead of wrangling IDCAMS.
- Built-in TN3270 terminal. Full PF, PA, and attention-key support, so the agent and a human can share the same session model.
The MCP foundation matters more than it first appears. Because the agent speaks the Model Context Protocol, Hopper isn't a walled garden — it can slot into the broader agent ecosystem the same way tools like E2B and OpenHands do, rather than reinventing orchestration.
The safety model
Execution is approval-gated: the agent pauses before it makes changes and waits for a human to confirm. In a production COBOL environment where a bad job submission has real-money consequences, this is the right default. It's not full autonomy, and you wouldn't want it to be.
Pricing Breakdown
Hopper has two tiers, and the gap between them is wide.
| Plan | Price | What you get |
|---|---|---|
| Hobby | Free | No credit card. macOS, Windows, and Linux clients. Connect to your own mainframe. All core Hopper features. |
| Enterprise | Custom | SAML SSO, MCP server access, admin and model controls, org-wide privacy controls, no model training on your data, on-prem / VPC deployment, SOC 2 and pen-test reports, priority support. |
The Hobby tier is genuinely free and includes the full feature set — that's unusually generous and a clear signal they want individual mainframe engineers evaluating it without a procurement cycle. The catch is the obvious one: you need a mainframe to point it at. There's no sandbox z/OS instance bundled in, so "free to try" only applies if you already have access to the platform.
Enterprise pricing is fully opaque. No published seat count, no module breakdown, no anchor number. For a regulated bank or government shop that's normal, but it does mean you can't estimate budget without a sales conversation — and the things you'd actually need for production (SSO, VPC deployment, the no-training guarantee, compliance reports) all live behind that wall.
Pros and Cons
What's good
- It targets a genuinely underserved market. Mainframe tooling has seen almost no AI investment. Millions of lines of COBOL still run critical systems, and the people maintaining them have had nothing comparable to a modern AI assistant.
- The free tier has no asterisks on features. No credit card, all core capabilities, your own mainframe — that's a real evaluation path, not a teaser.
- MCP-based architecture. It can integrate with the wider agent ecosystem instead of locking you into one vendor's runtime.
- Approval gating reduces blast radius. Pausing before changes is the correct posture for production z/OS, and it's built in rather than bolted on.
What's not
- Extremely niche. If your organization isn't running z/OS, none of this is relevant. This is not a general coding assistant — for that, look at Cursor or Replit.
- Opaque enterprise pricing. No published costs means no way to budget without sales involvement.
- The free tier still needs a mainframe. Casual evaluation is effectively impossible if you don't already have access to one, which most developers don't.
- Very early product. Limited public track record and almost no independent third-party reviews yet. You're an early adopter, with the risk that implies.
Who Is It For
This is a narrow tool, and that's fine — the question is just whether you're in the narrow band.
- Mainframe engineers and ops teams who spend real hours in SDSF triage, JCL debugging, and abend analysis. This is the target user, and the time saved on dump decoding alone could justify the trial.
- Banks, insurers, airlines, and government IT running z/OS who are starting modernization conversations and want AI tooling that meets the platform where it is rather than forcing a rewrite.
- Platform teams evaluating MCP-based agents who want to see how the protocol holds up against a genuinely hard, legacy target.
It is not for general application developers, anyone without mainframe access, or teams looking for their first AI coding assistant. Those builders are better served by mainstream tools.
Verdict
Hopper is a technically credible bet on a problem most of the industry has ignored: the COBOL still running critical banking and government systems has no modern AI tooling, and that gap is real. The product is early and the market is small, but the execution looks serious — approval-gated changes, MCP integration, structured abend diagnostics — rather than a thin demo.
If your team regularly wrestles with SDSF triage and JCL debugging, the free Hobby tier is worth a serious look — there's no credit card and no feature gate, so the only cost is your time. Connect it to a non-production LPAR and let it decode an abend or two before you judge it. For everyone outside the mainframe world, this is one to watch as the modernization wave builds, not one to adopt today.
Rating: 7.2 / 10 — strong concept and credible execution in a neglected space, held back by how early and how narrow it is. Try Hopper free if you have a mainframe to point it at.