z/OS
Learn more about features
z/OS Agentic AI · Batch + Interactive Workspace
AI Inference for z/OS —
Any Model. Zero Assembly.
Connect/AI
Native
z/OS Agentic
AI Platform
No JVM
No
Middleware
Zero
Application Code
Changes
3
Inference Model
Targets
SAF
Access
Control
TLS
Native z/OS
End to End
5
Output
Destinations
Enterprise AI on z/OS—scheduled in batch or interactive at your desk, secure and agentic.
Every Connect/MP license already includes governed MCP APIs for JES, FILE, and 3270, plus a chat-only inference workspace. Connect/AI adds what only Connect/AI delivers — agentic tool use in the workspace, and native batch agentic inference invoked straight from JCL.
Select mode ›
Embed AI into your z/OS batch jobs — on your terms.
Connect/AI lets you add powerful inference directly into existing batch workflows with zero application changes. Choose your model strategy:
- ›Mainframe-Resident Models — Run open-source models like IBM Granite via vLLM directly on your own z Systems LPAR for maximum security, lowest latency, and zero data movement.
- ›Local/Cloud Models — Access local and powerful external models when you need specialized capabilities or higher intelligence.
Inference Mode
A simple JCL step sends your data and receives a structured response. Conversation history is preserved automatically across steps, allowing each inference to build on the last — perfect for pipeline-style processing.
Agentic Mode
The same JCL step can launch a full reasoning agent. It plans, reads and analyzes z/OS resources (datasets, files, job output), observes what it finds, and iterates until the task is complete — entirely read-only, so nothing changes unattended. When the task calls for action, pair it with the Interactive Workspace, where every write is gated by your approval.
Projects
Projects aren’t limited to the Workspace — the same shared Markdown (.md) instructions carry over to batch, managed from Batch/CLI or the WebUI. Every JCL run starts fully briefed, so pipeline runs stay consistent without re-explaining setup step by step.
All execution runs under the job’s native z/OS identity with full SAF security, auditing, and compliance. Whether using local or cloud models, your data stays protected and governed by your existing policies.
Reliable, auditable AI inside your batch window — no middleware required.
JCL
CONNECT/AI
Batch
prompt / request
inference result
INFERENCE
MODEL
MODEL
Mainframe | Local | Cloud
read / observe
AGENTIC
multi-level iterative loop
multi-level iterative loop
reason
AGENT EXECUTION — READ-ONLY
READ
read z/OS resource
Files
JES
▶
OBSERVE
Structured result
REASONING
REASON
plan next step
Batch AI Inference
Enterprise-grade Security
Connect/AI runs as a native z/OS process, giving every prompt, model interaction, and dataset the same SAF access controls and audit policies as your most critical mainframe workloads. Traffic to external model providers is automatically secured with IBM System SSL encryption — keeping AI workflows safe, controlled, and fully accountable.
Why Connect/AI
Any Model. Any Data. Batch or Interactive—Natively on z/OS.
One Product, Not a Project
Complete integration layer — you bring the model, Connect/AI handles the rest
Every other mainframe AI capability — IBM or otherwise — is ultimately a collection of components that someone has to design, integrate, license, test, and operate. That's not a solution. That's a project. Connect/AI arrives as a complete, self-inclusive integration layer — no integration architecture to build, no middleware server, no specialist assembly required. Point it at your inference model — mainframe-resident (e.g., IBM Granite via vLLM on your own LPAR), your own local model, or a cloud provider like Claude, OpenAI, or Grok — add a JCL step or open the workspace, and run.
Native z/OS Agentic 3270 — Unlocked by Connect/AI
Plans and drives multi-step 3270 workflows under your approval — built on the MCP API already included with Connect/3270
Connect/AI mines live VTAM (CICS, IMS, OMEGAMON, etc.) screen content and reasons over it — using your choice of mainframe-resident, local, or external model such as Claude, OpenAI, or Grok — planning and driving multi-step workflows with your approval gating every write. It's built on the MCP API already included with Connect/3270, so that connectivity isn't the unlock; the agentic reasoning that plans and acts on it is what Connect/AI adds. Connect/3270 remains available on its own for browser-based human access to those same sessions.
Model Flexibility Without Lock-in
Local, mainframe-resident, or cloud inference — choose the right model for each workload
Support for local, mainframe-resident, and cloud-based inference models means organizations choose the right model for each workload — balancing cost, latency, data sovereignty, and capability — without rewriting their integration layer.
Data Sovereignty & Compliance by Design
Sensitive data never leaves the mainframe — satisfies GDPR, HIPAA, PCI-DSS, and SOX
Sensitive financial, healthcare, and government data never has to leave the mainframe. By enabling on-premises or mainframe-local inference, Connect/AI satisfies GDPR, HIPAA, PCI-DSS, and SOX requirements that cloud-only AI pipelines cannot.
Dramatic ROI on Existing Investment
Extends the value of existing MIPS, licenses, and skills — AI on top of proven infrastructure
Rather than expensive cloud migration or re-platforming, Connect/AI extends the value of existing MIPS, licenses, and skills. AI capabilities are layered on top of proven infrastructure — not instead of it.
Seamless Batch Workflow Integration
Fits inside existing z/OS job streams — full restart/recovery, SMF logging, operational visibility
No new runtime, no new ops model. Connect/AI fits inside existing z/OS job streams, meaning AI inference becomes just another step in a production batch process — with full restart/recovery, SMF logging, and operational visibility.
Modernization Without the Risk
Incremental AI adoption — legacy systems stay the system of record while gaining AI insight
For organizations on a modernization journey, Connect/AI provides an incremental path — extracting AI-driven insights from legacy systems while those systems remain the system of record. It bridges old and new rather than forcing a binary choice.
Empowers Mainframe Teams in the AI Era
A native, high-performance AI integration purpose-built for z/OS batch and web workloads
Connect/AI equips Systems Architects, AI/ML Engineers, Mainframe Architects, and z/OS Infrastructure Leaders with direct, native access to model-driven inference from z/OS batch and web workloads. Your model performs the inference — on Z or off-platform — while Connect/AI moves the data, returns the results, and enforces SAF authorization. Enterprise AI on mainframe data, with minimal disruption and full operational control.
External Agents and Applications
MCP APIs for JES, Files, and 3270 ship with every Connect/MP license — bring your own agent, or let Connect/AI run it natively
Governed MCP APIs for JES, Files, and 3270 come included with any Connect/MP component license — no Connect/AI purchase required. Bring your own agent: frameworks like LangChain, or integrations you build on Claude, OpenAI, or another model, can call these APIs directly to submit and track jobs, read and write datasets, and drive 3270 automation, with no terminal emulation or screen scraping required. Prefer not to build and host that integration yourself? Connect/AI runs the agent natively, on the mainframe — reasoning, planning, and acting under your approval, with no separate framework or middleware to operate. Connect/Web serves both paths as a native z/OS binary, with no JVM or middleware dependencies.
Click any card to expand details.
Resources
Connect/MP › Connect/AI
Features
AI Inference
Per-step model selection
Each JCL step independently specifies its own inference model
Connect/AI lets every step in a batch job choose the model that best fits its task — a lightweight local model for classification, a cloud model for narrative generation, a mainframe-hosted model where data residency is required. Cost, capability, and compliance requirements are addressed per step without restructuring the job.
JCL-native invocation
Direct from batch JCL — no middleware, no migration required
Connect/AI is invoked like any other z/OS utility: an EXEC PGM step in JCL with DD statements for the prompt template, input dataset, and output target. No new languages, no new toolchain, no external middleware. Existing JCL skills apply immediately.
Multi-model support
Local, mainframe, and cloud inference — all from the same JCL framework
Connect/AI reaches your own on-premises models, z/OS-hosted inference engines, and major cloud providers (Claude, OpenAI, Grok). Different steps in the same job can use different model targets, mixing local and cloud as workload needs dictate.
Projects with shared MD instructions
Faster, more consistent batch results — without paying setup costs every run
Long-running work — a migration, a modernization effort, a recurring investigation — lives in a Project. A single Markdown (.md) file captures your standing instructions and environment context once, and every JCL run in that Project draws on it automatically, managed from Batch/CLI or the WebUI. The payoff: faster time-to-value on every run, consistent outcomes job after job, and institutional knowledge that compounds instead of being re-explained from scratch each time.
Interactive Workspace
Real-time interactive inference
Prompt models and explore mainframe data live from a browser workspace
The Connect/AI workspace brings real-time AI interaction to z/OS. Analysts and developers prompt inference models, test prompt templates, and explore mainframe data interactively — without waiting for a batch window. The same models available to scheduled jobs are available live at your desk.
Agentic multi-step workflows
Models that plan and act across steps, not single-shot inference
Connect/AI is agentic. The workspace runs multi-step workflows where the model reasons, calls tools, and acts on intermediate results — chaining inference, data retrieval, and follow-up actions to complete a task rather than returning a single answer.
Projects with shared MD instructions
Faster, more consistent agentic results — without paying setup costs every run
Long-running work — a migration, a modernization effort, a recurring investigation — lives in a Project. A single Markdown (.md) file captures your standing instructions and environment context once (e.g. "assume IBM Enterprise COBOL," "this is for LPAR SYSA"), and every agentic run in that Project draws on it automatically. The payoff: faster time-to-value on every task, consistent outcomes run after run, and institutional knowledge that compounds instead of being re-explained from scratch each time.
Agentic 3270 — unlocked with Connect/3270
Drive live 3270 sessions from the interactive workspace
Paired with Connect/3270, the workspace reaches beyond datasets and files into your 3270 applications. Connect/3270 mines live VTAM (CICS, IMS, OMEGAMON, etc.) screen content and structures it; Connect/AI's agentic workflows — powered by your choice of mainframe-resident, local, or external model such as Claude, OpenAI, or Grok — reason over that data and drive multi-step 3270 interactions, with your approval gating every write.
Browser-based, any device
Secure z/OS-native access with no desktop client to install
The workspace is delivered entirely in the browser — no emulator, no desktop client, no local install. Users connect securely from any device, and all processing stays on z/OS.
Powered by Connect/Web
Served by the native z/OS HTTP engine — zero JVM, same SAF security
The interactive workspace runs on Connect/Web's native z/OS HTTP engine. There is no JVM layer and no separate web server to manage. Interactive AI inherits the same SAF access control and data residency guarantees as batch inference, with IBM System SSL automatically securing any external model connections.
Output & Routing
Dataset & file output
Write inference results to z/OS datasets for downstream batch steps
Inference output writes to sequential datasets, PDS members, or VSAM files — exactly like any other z/OS batch utility. Downstream job steps can read and process AI output without format conversion or middleware.
Formatted report delivery
AI-generated output to SYSOUT — printable and archivable
Connect/AI can write formatted reports to SYSOUT, making AI inference output available in the standard batch report stream — ready for existing distribution, archiving, or printing workflows.
Connect/Hybrid routing
Pipe inference output into distributed applications
Inference results can be routed through Connect/Hybrid into distributed systems — CRM platforms, cloud queues, REST APIs, or messaging systems — without changes to the z/OS batch job or the receiving application.
Email & FTP delivery
Route AI-generated results direct to SMTP or FTP targets
Connect/AI can deliver inference output directly to email recipients via SMTP, or push results to remote systems via FTP — making AI-generated reports and summaries available outside the mainframe environment without additional tooling.
Security & Compliance
SAF access control
Resource-level authorisation using existing mainframe security policy
Connect/AI is fully integrated with SAF. Every inference job runs under the submitting user's SAF identity, with resource-level access control on model targets, prompt templates, and output datasets. No new security policy is required — AI inference inherits the controls already in place.
System SSL encryption
Cloud model connections automatically secured with IBM System SSL
Connections to cloud model providers that require encryption — such as Claude and Grok — are automatically secured using IBM System SSL, with no third-party TLS library required. Local and mainframe-resident model traffic stays on-platform and never leaves your system.
Data residency control
Keep sensitive data on-premises with local or mainframe models
Connect/AI supports local and mainframe-hosted inference models for workloads where data cannot leave the on-premises environment. Compliance-sensitive data — policyholder records, government datasets, financial transactions — stays on z/OS throughout the inference pipeline.
Connect/MP › Connect/Web › Connect/AI
The Web AI Workspace — Your Native z/OS AI Console
A powerful, browser-based AI console that brings true agentic intelligence directly to your mainframe environment. Unlike simple chat interfaces, it enables multi-step iterative reasoning and agentic execution — with your approval gating every write — all running natively on z/OS with full security and control.
Key Capabilities
Scroll to see all columns
| Function | Description | Value for Mainframe Users |
|---|---|---|
| Multi-Step Iterative Reasoning | AI can break down complex tasks, plan steps, execute, observe results, and loop until the goal is achieved | True agentic workflows from a single prompt |
| Tool Use & Integration | Connects to external tools (files, JES, 3270, databases, etc.) in real time | Direct action on z/OS resources |
| Live Screen Interaction | Reads and navigates 3270 screens, extracts the data | Agentic 3270 Inference |
| Real-time Observation & Feedback | Shows reasoning steps, intermediate results, and final output | Transparency and control |
| Interactive Prompting | Users can guide, correct, or continue the agent mid-process | Human-in-the-loop capability |
| Conversation Memory | Maintains context across multiple steps and sessions | Complex, long-running tasks |
| Projects & Shared Context | A named Project bundles a shared Markdown (.md) instructions file across many chats and runs | Consistent results without re-explaining setup every time |
| Structured Output | Returns clean, structured data, summaries, reports, or actions | Easy integration with downstream processes |
| Model Flexibility | Switch between mainframe, local, or cloud models on the fly | Balance cost, security, and capability |
| Audit & Security | All actions run under the user’s SAF identity with full logging | Enterprise compliance |
| Browser-Based Access | No installation — works from any authorized desktop | High productivity for analysts & developers |
This console allows analysts, developers, and operations teams to perform complex, multi-iterative tasks using a single prompt — combining reasoning, live 3270 interaction, file/JES operations, and more — without leaving the browser.
This site uses cookies. By continuing to browse our site you are agreeing to our use of cookies.