AGENTIC WORKFLOW ORCHESTRATOR

Fig. 0 — what it is

Wire inputs, AI steps and actions into a graph

Several inputs can feed one graph, each running only the part it reaches — event-driven, batched, scheduled or on demand. Runs on your machine, with any model, and tells you plainly when something breaks.

Download AWO — free See how it works

Free tier, no account, no telemetry. Node.js 22+.

stock-watch.yamltwo inputs, one graph
sources:
  - id: chat
    trigger:
      type: discord
      botToken: ${DISCORD_BOT_TOKEN}
      channelIds: ["1189…"]
    # 25 messages, one AI call
    deliver: { mode: batch, maxWaitSeconds: 300 }
  - id: hourly
    trigger: { type: schedule, everySeconds: 3600 }

nodes:
  - { id: read, type: reason, prompt: "Which stocks?" }
edges:
  - { from: chat,   to: read }
  - { from: hourly, to: read }

The problem with one-thing-after-another

Most automation tools give you a list of steps that runs top to bottom. That is fine until the work is real, and then three things break at once.

Two things start the same job

A message arrives and a schedule fires. With a list you build two near-identical workflows and watch them drift apart, or one that does both jobs on every event and bills you for the half you did not need.

The AI's answer should decide what happens next

"Is this urgent?" ought to pick one of three branches. A list can only run the next step, so the branching ends up inside a prompt, where you cannot see it, test it, or find out why it chose wrong.

A busy channel costs a fortune

One run per message is fine at ten messages a day and ruinous at three hundred an hour — every message its own model call, most of them saying nothing worth a call.

How AWO works

You draw a graph. AWO runs the part of it that each event actually reaches, and keeps a record of what happened and why.

Wire up what happens

Add inputs, AI steps, conditions and actions from the palette, and say what runs next. The canvas is a plain YAML file on your disk — the same document the command line runs.

Press Test run

The graph runs for real with a stand-in event. Every step shows what it was given, what it produced and why it ran — including the steps that were skipped and the reason.

Set it watching

It runs on its own from then on: immediately on an event, batched over a window, on a schedule, or when you say so.

Every feature, explained in full →

Fig. 2 — one graph, three inputs

Which part of the graph runs — and which part doesn’t. Pick an input to take over.

Nothing has run yet — pick an input above.

What you compose it from

Nine primitives. Everything else is how you wire them.

Inputs

Discord, Telegram, webhooks, live sockets, page changes, schedules, manual

Ask the AI

Any model, per workflow or per step. Reply as text or parsed JSON

Pull out data

A field from an event, or text matching a CSS selector

Only continue if…

Gate the branch on a real value

For each item

Fan out over a list, with ceilings on breadth and spend

Bring together

Merge branches and read what actually arrived

Send an alert

Discord, Slack, Telegram, email, database, any webhook

Run a command

Tests, builds, scripts. Off until you allow-list it

AI writes code

A coding agent in a sandbox, opening a PR. Off by default — and the basis of Autopilot

What people build with it

Each of these ships inside AWO as a recipe: open it, fill in two or three fields, press Test run.

Watch a page and tell me what changed

Checks on a schedule, asks the AI whether the change was worth a message, and only then alerts you.

scheduleask the AIonly if…alert

Triage a busy chat channel

Reads a batch of messages, classifies what is being asked, and routes each kind down its own branch.

discord · batchedclassifyroute

Reproduce → fix → verify

Turns a bug report into a failing test, fixes until it passes, and refuses to open the pull request if it does not.

reportfailing testfixgate

All eleven recipes, with the steps they wire →

And for software engineering work: Project Autopilot

Point AWO at a repository and give it a goal. It drafts a plan of phases with dependencies, shows you the plan before anything runs, then works it phase by phase — opening one pull request per phase with your test suite as the gate.

A plan you can edit

The goal is decomposed into phases. Add, reorder, split or delete them before a line of code is written — the decomposition is a document, not a hidden step inside a prompt.

Five cycles per phase

Researcher and planner investigate — both read-only, and no config can make them writable. Builder writes and opens the PR. Reviewer and tester gate it: fail, and there is no pull request.

Done means done

A run is complete when every phase is done with a landed PR — a fact about saved state, never a model announcing it has finished.

How Autopilot works, in full →

What batching saves you

A busy channel is where one run per message stops being affordable. Put your own numbers in — a model call is the unit of cost, so the saving is just how many calls stop happening.

One run per message

AI calls / day

Batched

AI calls / day

Fewer calls

at the same coverage

Saved / day

at $0.002 per call

Batching never runs less often than its own timer, so a quiet channel still delivers on the window. The arithmetic above accounts for that.

What a run tells you

Every step keeps what it was given, what it produced, how long it took, and why it ran. Open one.

When it goes wrong

The same failure, reported two ways. The first is what the machine knows. The second is what you need.

The Discord message could not be sent because it has no message text.

Open this step and fill in message text — either type it, or reference an earlier step’s output like {{ ai.output }}.

And most never reach a run: AWO checks what it can while you are still configuring, so the mistake is caught before the model is ever called.

The thing itself

The builder, as it actually looks. Steps carry a coloured seam by kind, the canvas is the same document the CLI runs, and a test run lights up every step it touched.

The AWO Studio: step types down the left, a four-step workflow on the graph-paper canvas — watch a page, ask the AI, only continue if, send an alert — and the selected step's settings on the right.
AWO Studio · a page-watch workflow, built from a recipe in under a minute

How it compares

Only things that can be checked. No opinions about anyone else’s product.

AWOHosted automation toolsWriting it yourself
Runs onYour machineTheir serversYours
Your data goesOnly where you send itThrough themOnly where you send it
ModelAny — local or cloudUsually theirsAny
Several inputs, one graphYesVariesIf you build it
Cost at 10k eventsFlatPer task or per runYour time
Works offlineYes, licence includedNoYes
You maintainA workflowNothingAll of it

Why a graph

Most automation tools give you a list of steps that runs top to bottom. That works right up until real work shows up:

A list cannot express any of those honestly. A graph can, and AWO’s engine executes it directly: an event arrives, AWO works out which steps it affects, gathers their inputs, runs them, records the result and carries on.

What it will not do

No telemetry

AWO makes no call home, ever. Not for licensing, not for usage, not for crash reports. The only requests are the ones your workflow makes.

No account to run it

Download, unpack, run. An account here exists only so you can read back a licence key you bought.

Off by default

Shell commands, filesystem access, git writes and code-writing agents each need a separate, explicit opt-in. A workflow file can never enable them on its own.

Your workflows are files

Plain YAML on your disk, readable and version-controllable. Stop paying and they keep working within the free limits.

What it costs

The free tier is a real product, not a demo. You pay at the point you want a second input feeding one graph, or a second workflow running.

Free

$0 / forever

No account, no card, no time limit.

  • One workflow, one input
  • Every node type, any model
  • Full run history and debugging
Download free

Questions people actually ask

Do I need to be a developer?

No, and the recipes are there for exactly that reason: pick one, fill in two or three fields, press Test run. You never see YAML unless you want to.

You do need to install Node.js once and choose a model. Both are one-time, and the download page walks through them.

Where does it run?

On your machine. The Studio is a local web app at 127.0.0.1, nothing is hosted, and there is no account needed to run it. If you close the laptop, the workflow stops — put it on a server you control if you want it always on.

Which AI models does it work with?

Any of them. A local Ollama costs nothing and works offline; OpenAI, Anthropic and your own endpoint work too. You can set a model per workflow, or per step — a cheap one for classification, a strong one for the summary that gets published.

AWO ships no model and no keys, so the choice and the bill are yours.

What does it cost to run?

AWO itself is free for one workflow. The running cost is your model calls, and batching is the lever that controls them — the calculator above takes your own numbers.

Is my data sent anywhere?

No. AWO makes no call home — not for licensing, not for usage, not for crash reports. The only network requests are the ones your own workflow makes, to the places you configured. The full account is here.

How is this different from Zapier, n8n or Make?

Three things, all checkable: it runs on your machine rather than their servers, it uses any model including a local one, and several inputs can feed one graph with each running only the part it reaches. The comparison table sticks to facts and offers no opinion about anyone else's product.

What happens if I stop paying?

Your workflows stay on your disk as plain YAML and keep working within the free limits. Nothing is deleted and nothing is held hostage. A lapsed subscription also gets 14 days at the paid tier first, with a notice on screen, rather than stopping at midnight.

Is it open source?

No — AWO is proprietary, licensed and not sold, and the download contains the compiled product. Your workflows are a different matter: plain files that belong entirely to you.

Build your first workflow tonight

Free tier, no account, nothing sent anywhere. Unpack it and the Studio opens — a recipe gets you running in about a minute.