Write one workflow.
Run it as a service.

ZipperGen is a Python framework for workflows with LLM agents, people, and services. Write who does what and who makes each decision. ZipperGen runs the participants, saves their progress, and keeps track of human approvals. Use one CLI to configure, deploy, and inspect the workflow on your machine or server.

Try the interactive demo Get started View on GitHub

Start with an approval workflow

This workflow waits for a message, asks a model to draft a reply, and asks you before sending it. Then it waits for the next message. 

@workflow
def email_approval() -> int:
    Mailbox: item = next_unread_message(processed)
    while (item is not None) @ Mailbox:
        Mailbox: message = message_text(item)
        Mailbox(message) >> Writer(message)
        Writer: draft = draft_reply(message)
        Writer(draft) >> Mailbox(draft)
        Mailbox: approved = approve_reply(draft)
        if approved @ Mailbox:
            Mailbox: handled = send_reply(draft, handled)
        else:
            Mailbox: handled = discard(handled)
        Mailbox: processed = complete_message(item, processed)
        Mailbox: item = next_unread_message(processed)
    return handled @ Mailbox

@ Mailbox says who owns the decision. The actions define how messages are read, replies are drafted, and approval is requested. The tutorial starts with a local folder and approval in the terminal.

Run it and see what is happening

Before deployment, configure the model and any external services the workflow needs. Choose how you want to receive approval requests, for example through Telegram. Then check and deploy the project: 

zg validate
zg check --strict
zg deploy

ZipperGen starts the service and saves its progress. Pending approvals stay available across a restart. From the same project directory, you can see what needs attention and how the service is running: 

zg deploy tasks
zg deploy status
zg deploy logs

Answer approval requests in the configured chat or use zg deploy approve. The deployment guide covers setup and running the service on macOS or Linux.

Work in your usual editor

A project is an ordinary directory with a Python workflow, a plain-text specification, and a small configuration file. Write the workflow yourself or ask a coding agent such as Claude Code or Codex to help. zippergen skill provides the authoring instructions. You and the agent use the same CLI. The code stays in files you can read, review, and change.

One workflow for all participants

ZipperGen derives a program for each participant from the global workflow. Each program contains only the work and communication that participant needs. In the example above, the Writer has no approval branch because it has no work in either outcome.

This approach is called choreographic programming. For well-formed workflows covered by the formal model, the generated programs cannot deadlock. The core theorem is machine-checked in Lean 4. CPL guards can also check earlier events and values that are causally visible at a decision. The README explains the scope of these guarantees.

Read the paper Core theorem checked in Lean 4

Bollig, Függer, Nowak. Provable Coordination for LLM Agents via Message Sequence Charts. arXiv:2604.17612. ISoLA 2026.

Bollig. Deadlock-Free Parallel Regions for Projected Workflows. EXPRESS/SOS 2026.

Bollig. Causal Past Logic for Runtime Verification of Distributed LLM Agent Workflows. arXiv:2605.20923. ICFEM 2026.