It scales — in capacity and in change
From one JVM to a cluster with the same definition; millions of waiting processes cost storage, not threads. And adding a step, a branch or a whole new process stays as easy on day 500 as on day one.
The video above is chapter 1 of a series of short chapters, one aspect at a time — every screen is the real product running. Each guide page embeds its chapter. Full playlist.
A business process starts as a few calls in a row — and then production adds the rest: undoing what succeeded when a later step fails, retries that must not book twice, a pod that dies halfway, people who answer two days later, code that changes while runs are in flight, and someone asking “where is booking 123?”. Written by hand, that is most of the code, and almost none of it is about your business.
All of it is yours to write, test, review and fix — in every service.
You write the tip. The rest is declared — or simply there.
It scales — in capacity and in change
From one JVM to a cluster with the same definition; millions of waiting processes cost storage, not threads. And adding a step, a branch or a whole new process stays as easy on day 500 as on day one.
You write less code
The process is a short, reviewable declaration. What remains in code is the business work inside each step — not state tables, sweepers, outboxes and locks.
The hard parts come for free
Retries, timeouts, sagas and compensation, parallel fan-out, timers, human tasks, crash recovery and versioning — declared in one line, solved once for every process.
It saves time — at every stage
Not just the first version: each change is one validated block, a definition change ships by merging to Git — no service redeploy — and tests run in memory with simulated time.
It can be operated
Every run is a live graph you can search, inspect, retry from a step, migrate or repair in bulk — audited, and from day one rather than after the first incident.
One screen per run
Its graph, steps, variables, errors and history, together — instead of correlating logs across services to rebuild what happened.
And when hand-written code is the better choice, we say that too: Why a process engine?
Distributed by design
Multiple orchestrator instances coordinate safely: events are keyed by process onto Kafka partitions, so each process has a single owner, and state is reconciled through the transactional outbox pattern. Scale from a single JVM to a multi-pod Kubernetes cluster without changing a line of business code.
Event-driven, not polling
Every state transition is triggered by a domain event — stored in the outbox table and relayed to the appropriate handler. No polling loops that grow with your data.
Developer-first DSL
Workflow definitions are JSON files — human-readable, version-controlled, reviewable in a PR. Retries, timeouts, compensation (saga), parallel execution, sub-processes, conditional branching (JEXL), and human tasks in a single flat file.
Three deployment modes
Run fully in-process for tests, single-node with PostgreSQL, or fully distributed with Kafka + PostgreSQL — no code changes required.
Built-in forms engine
User-task steps pause the workflow until a human submits a form. Forms are defined in JSON, stored in version control, and served dynamically to any front-end.
Native AI via MCP
The first workflow engine with native Model Context Protocol (MCP) integration. Any MCP-compatible AI client — Claude Desktop, a chatbot, an internal copilot — can operate the engine in natural language.