Security

Designed for people whose job is to say no

MainframeLore was built for banks, insurers, and government mainframe shops — the most security-conscious buyers there are. Here's exactly what happens to your source code, what deployment options exist, and what we don't do.

The core guarantee

No mainframe access, ever — under any tier

How analysis actually happens

You export your COBOL, copybooks, and JCL — the same PDS export you already produce for source control, audits, or vendor reviews. That export is the only input. There is no agent installed on z/OS, no CICS transaction, no network path opened into your mainframe, and no credentials of any kind requested for your systems. The tool reads files; it does not connect to anything.

Deployment options

Pick the posture that matches your risk tolerance

TierWhere processing happensWhere your source lives
AssessHosted, one-time runDeleted after the report is delivered
TeamHosted, continuousRetained only to re-run docs on new drops; deletable on request
EnterpriseInside your own network (container)Never leaves your infrastructure

AI and your data

What touches an LLM, and what doesn't

Common questions

What security teams usually ask

Do you store our source code?

On the Assess tier, no — it's deleted once your report is delivered. On Team, only what's needed to re-run documentation on new source drops, and it's deletable on request. On Enterprise, never — everything runs inside your own network.

Does the AI model get trained on our code?

No. Requests to Anthropic's API are not used to train models per Anthropic's API terms. If your policy requires it, use the Enterprise tier with your own LLM deployment for a fully self-contained pipeline.

Can this run fully offline / air-gapped?

Yes, on the Enterprise tier — the parsing and analysis pipeline runs inside your network with no external calls, and either a self-hosted LLM or the deterministic analysis-only mode replaces the hosted AI step.

What if we can't share real production code at all?

Start with a sanitized or representative program — the parser works identically on any valid COBOL/JCL. Many customers validate output quality on a non-sensitive program before sending anything production-facing.

Questions we haven't answered here?

Ask directly