# AI agents for the mainframe.

Zcrafter builds collaborative bots and the RL environments that train them.

[Request access](https://cal.com/zcrafter/zcrafter-demo)

*On the web page, a working demo of the Cobots app is shown here.*

## Backed by

- Claude for Startups
- Cloudflare for Startups
- 1871 Chicago
- The Linux Foundation
- Open Mainframe Project

## Our software

### Cobots

A team of agents that works on your mainframe. Every change waits for your approval.

### Bench

Reinforcement-learning (RL) environments where agents practise mainframe work and are scored on the result.

## Cobots: a team of agents on your mainframe.

### Reads run now

Agents list datasets, browse members and read spool on their own. Reading changes nothing.

![Cobots: the z/OS Systems agent has listed six datasets in the chat, with the mainframe screen and the two commands it ran beside it](https://zcrafter.com/assets/images/cobots/reads-2400.webp)

### Changes wait for you

Edits, job submits and console commands stop at an approval card that shows the exact change.

![Cobots: an approval card from the COBOL Dev agent showing the one line it wants to add to a member, where the old version will be kept, and the buttons Allow once, Edit and Deny](https://zcrafter.com/assets/images/cobots/approval-2400.webp)

### Every screen, in the open

Each agent has a live 3270 screen and the list of commands it ran, in order.

![Cobots: the Batch Ops agent’s mainframe screen opened large, showing the job log of a job that ended with condition code 0000](https://zcrafter.com/assets/images/cobots/screen-2400.webp)

### Yours to run

A desktop app for macOS and Windows. Passwords stay in your keychain. Bring your own model.

![Cobots: the Models settings, with OpenAI Codex and Anthropic connected, other providers and a self-hosted server offered, and a list of Claude models to choose from](https://zcrafter.com/assets/images/cobots/models-2400.webp)

[Request access](https://cal.com/zcrafter/zcrafter-demo)

## Bench: RL environments where agents learn the job.

### Task environments

Isolated, resettable tasks with COBOL, JCL, datasets, jobs and failures. Two attempts face the same conditions.

**An example task: payroll abend recovery**

Start: **PAYRUN** `ABEND S0C7` → The agent: **Finds the fault** and fixes it → Goal: **PAYRUN** `CC 0000`

The system is reset before every attempt.

### Controlled tools

TN3270, MCP and APIs for inspecting state and acting. The tools enforce permissions, not the model.

**What the agent may use**

**Allowed**

- Read datasets, spool, jobs
- Edit `PAY.TEST.**`
- Submit `PAY.TEST.JCL`

**Denied**

- `PAY.PROD.**`
- Console commands
- Security changes

### Deterministic rewards

Correctness, safety and efficiency, checked on the system. The same outcome always scores the same.

**How an attempt is scored**

- Compiled: PAYCALC builds
- Completed: PAYRUN ends CC 0000
- Correct: the report matches
- Contained: protected datasets unchanged

Reward: 4 of 4

### Agent trajectories

Observations, tool calls, state changes and outcomes, recorded. How a result was reached matters as much as the result.

**One attempt, as recorded**

1. observe: PAYRUN abends, S0C7
2. tool: Read PAYCALC
3. tool: Read TIMECARD
4. change: Edit PAYCALC
5. tool: Submit PAYRUN
6. outcome: CC 0000

[Contact sales](https://cal.com/zcrafter/zcrafter-demo)

## Bridge: an open-source MCP server for z/OS.

### One line to install

Use the AI agent you already have. Nothing is installed on the mainframe.

**How it connects**

Your AI agent: **Claude, Codex, Cursor** or any MCP client → On your machine: **Bridge** 48 tools, one gate → Your mainframe: **z/OS** through z/OSMF

Install: `uv tool install zcrafter-bridge`

### 48 tools

Datasets, members, jobs, spool, z/OS UNIX files, TSO and the console.

**What an agent can work with**

- **11** Datasets
- **7** Members
- **15** Jobs and spool
- **7** z/OS UNIX files
- **3** TSO and console
- **5** System

23 read. 25 change the system, and wait for your yes.

### You approve every change

Reading is free. A change runs only if it is a step in a plan you approved.

**What you are asked**

Approve this plan?

1. `member.patch` PAY.TEST.COBOL(PAYCALC)
2. `job.submit` PAY.TEST.JCL(PAYRUN)

Buttons: Approve, Reject.

Only these two steps can run. The agent cannot approve its own plan.

### Open source

MIT licensed. Passwords stay in your keychain and never reach the model.

**Where to get it**

- GitHub: [`ZcrafterAI/Bridge`](https://github.com/ZcrafterAI/Bridge)
- PyPI: [`zcrafter-bridge`](https://pypi.org/project/zcrafter-bridge/)
- MCP Registry: `io.github.ZcrafterAI/bridge`
- License: `MIT`

[Get Bridge on GitHub](https://github.com/ZcrafterAI/Bridge)

## Technical overview

How Cobots and Bench are built, hosted and secured.

PDF, 4 pages. November 2026.

[Read the overview](https://zcrafter.com/assets/docs/zcrafter-technical-overview.pdf)

## Research

Design notes on RL environments, evaluation, and mainframe interaction.

### Designing rewards for batch-job recovery

What should count as success when a job completes but its output is wrong?

#### The evaluation problem

A recovery task can pass one checkpoint and still fail the overall objective. A program may compile, a job may complete, and the resulting records may still be incorrect. A useful evaluation needs an explicit definition of the expected final state.

#### Separate the checks

Our proposed evaluation separates compilation, job completion, expected output, and protected-resource integrity. Each check answers a different question. Keeping their results separate would make it possible to distinguish a failed repair from an unintended change elsewhere in the environment.

#### Make the starting state reproducible

The task definition should identify the source, job configuration, input datasets, and expected output. A reset should restore those inputs before another attempt. Otherwise, an earlier attempt could change the conditions under which a later agent is scored.

#### Avoid rewarding shortcuts

A combined reward needs to account for correctness and the boundaries of the task. Completing a job after bypassing a required check or modifying a protected dataset should not be treated as a successful recovery. The open question is how to combine these constraints without obscuring the reason for failure.

### Representing datasets in disposable environments

Records, access boundaries, and reproducible starting states.

#### The representation problem

An agent environment needs to preserve the dataset properties that affect the task. Treating every dataset as an ordinary text file may lose information that matters to program behavior, permissions, or the interpretation of an output record.

#### Start with a task fixture

A proposed fixture would identify the source, input and output datasets, record structure, and permitted actions. The goal is to specify the information needed to complete and verify a task before adding more simulation detail.

#### Reset more than the source

Reset semantics should cover input records, generated output, job configuration, and relevant metadata. Restoring the program alone would not be sufficient if a previous attempt had changed the data it reads or the resources it can access.

#### Make boundaries observable

The environment should distinguish resources the agent may modify from those it must leave unchanged. A verifier could compare the final state against those boundaries, while the trajectory records attempted actions that did not result in a successful modification.

#### The open question

The minimum useful fidelity remains unresolved. We need to establish which dataset semantics change agent behavior and which can be simplified without changing the task. This is a design proposal, not a validated simulator or an assertion of compatibility with a licensed mainframe system.

### TN3270 as an agent interaction surface

Representing fields, cursor position, and screen state in an agent observation.

#### Beyond a text transcript

Field-based interaction introduces structure that a plain transcript can lose. Screen contents, field attributes, cursor position, and allowed actions can all affect what an agent should do next.

#### A proposed observation

One representation could pair visible screen contents with a structured description of fields and the current cursor position. The observation should make the available actions clear without assuming that every piece of displayed text is an editable input.

#### Record the interaction

A trajectory should connect the observation to the action taken and the state that follows. This would help explain whether a failure came from an incorrect diagnosis, the wrong field selection, or an unexpected screen transition.

#### Verify the operational result

Reaching a particular screen does not by itself establish that a task succeeded. The proposed verifier should check the relevant job and dataset state independently of the agent’s account of what it did.

#### Validation still needed

This proposed interface has not yet been validated. Evaluation would need representative workflows and a clear distinction between simulated terminal behavior and access to licensed or customer-controlled systems.

## Links

- [Request access to Cobots](https://cal.com/zcrafter/zcrafter-demo)
- [Technical overview (PDF)](https://zcrafter.com/assets/docs/zcrafter-technical-overview.pdf)
- [Bridge on GitHub](https://github.com/ZcrafterAI/Bridge)
- [Zcrafter on GitHub](https://github.com/ZcrafterAI)
- [Zcrafter on LinkedIn](https://www.linkedin.com/company/zcrafter)
- [Zcrafter on X](https://x.com/Zcrafter)
- [Founders: LinkedIn 1](https://www.linkedin.com/in/victor-taiwo21)
- [Founders: LinkedIn 2](https://www.linkedin.com/in/oyewaleadeoye)
- [Privacy](https://zcrafter.com/privacy.md)

---

This is https://zcrafter.com/ as Markdown, written from the same source as the web page. An index of the site for AI agents is at https://zcrafter.com/llms.txt.
