AI workflow generator
Describe it. Get a real pipeline.
Most AI automation stops at suggesting steps. Describe a use case to the Zenvara workflow engine and it drafts the actual pipeline, as typed YAML you can read, review and keep in Git.
How it works
Three steps, no canvas to draw.
You stay in a conversation. The engine does the writing.
- 1
Say what you want automated
Plain language. "When an invoice lands in the shared mailbox, pull the line items, check them against the ledger, and open a ticket if they disagree."
- 2
Zenvara drafts the pipeline
The assistant writes the flow through the platform's own entity tools. You get typed YAML, not a screenshot of a graph, and the DAG renders so you can see the shape of what it built.
- 3
Review it, then run it
It compiles and validates before anything executes. You read the diff, change what you want, and run it. The YAML is yours: version it, review it, deploy it like any other code.
What comes with the engine
Describe it once. Everything after that is the engine’s job.
Writing the pipeline, proving it before it runs, keeping types intact across every step, undoing its own writes when something fails, and handing an agent the whole platform to operate. All of it comes with the engine.
Validated before it runs
The flow is compiled into a typed DAG and checked first. Unknown steps, bad references and type mismatches surface before execution, not halfway through a run against production data.
It undoes its own writes
Write connectors carry a compensating action. If step six fails, the writes from steps one to five are unwound. You are not left reconciling half-finished work by hand.
Types survive the step boundary
Expressions evaluate with real types rather than string coercion, so a number stays a number on its way from one step to the next.
The agent can operate the platform
54 MCP tools cover entities, runs, deployment, secrets and introspection. An agent can inspect a run and fix the flow, not just trigger it as a black box.
Once it is built, it runs for free
The AI helps you build the flow; it is not needed to run it. A published pipeline is a compiled program, so the thousandth run costs what the first did — nothing — and returns the same answer. No per-run model bill, and no model drift in your production path.
Your people and the agents share one room
Chat rooms are first-class members, not a side panel. Several agents can join under their own labels, message and direct-message each other, and wait on each other’s turns, while the people in the room read along live and the history persists. Enabled per deployment.
Three ways in
Talk to it, build it yourself, or bring your own assistant. Same engine, same room.
You are not choosing a product tier. You are choosing how much of the work you want to do by hand on any given day.
Basic — describe it
A conversation. Say what you want automated in plain language; the assistant drafts the flow as typed YAML and publishes it, then takes you to the graph so you can read what it built. Nothing reaches a deployed flow because a model proposed it. From the next release it does everything Pro mode does: every kind of flow, built, changed and run from the conversation.
Pro — build it
The full Studio: eight editor surfaces, the YAML, the run history, the secrets, the environments. Everything the assistant can do, you can do directly — and the two are the same live document, so you can start in chat and finish by hand.
Your assistant — drive it from there
Connect Claude, or any assistant that speaks MCP, to your Zenvara and work from the chat you already use: it can build, run and inspect flows there. Listings in the Claude and ChatGPT app directories are on the way.
What it costs
Cheap to build with. Free to run.
Most AI tools charge you every time the automation runs. Zenvara spends AI while you build, where it helps, and keeps it out of the run unless you put it there.
Cheapest model first
The assistant starts on the cheapest model of the provider you chose, such as Claude Haiku, and only moves up to a stronger one, such as Sonnet or Opus, when the task needs it. On by default for Anthropic, OpenAI and Google.
No tokens per run
A published flow runs without the AI that wrote it, so drafting it costs nothing per run. A flow only calls a model where you add an AI step yourself.
Or no API bill at all
Point Zenvara at a model you host, through Ollama, LM Studio, vLLM or llama.cpp, and building costs nothing but your own hardware.
Before you plan around it
The edges, stated up front.
Zenvara is in public beta and under active development, with at least one release a month. These are the limits worth knowing before you plan around it.
- The canvas is a viewer, not an editor
- The DAG shows you a compiled flow so you can read and inspect it. You do not draw a flow by dragging nodes. Authoring happens in chat, from a template, or in the YAML directly.
- You review before it ships
- Generated output is a draft by design. Nothing reaches a deployed flow because a model proposed it — you read it and decide, which is the point.
See it on your own use case
Bring a workflow. We'll draft it live.
Describe something you automate today, or wish you could. We'll build it in the session and you keep the YAML either way.