PipeDuck · running in production

PipeDuck: Automation & AI you can actually put in production.

A script when the rules are known, a model when they are not, a person when it matters - wired together on one canvas, on your own infrastructure.

  • A library of scripts and connectors, ready to use. Build your own for anything specific to your business.
  • AI where it helps, fixed logic where it counts. You pay only for the AI you use.
  • Nothing runs unapproved. A model drafts, a person signs off, every version is kept.
  • When something breaks, you get the trace and the retry. Watched live on a dashboard, not buried in a log.
  • Cost control, built into how it runs. The cheapest option for you, every time - not the flashiest.
  • Every dead end has an AI way through. Scripting, building, debugging a run - help is one click away, billed only for that click.
Client mail → answer sent 7 nodes
yes no TRIGGER Client mail arrives SCRIPT Read the thread CONDITION Needs an answer? AI MODEL Draft reply SCRIPT File & stop APPROVAL You sign it off ACTION Send the answer
Start

Choose per step. Not per platform.

A process is not one kind of work - so you are not made to pick one kind of tool. Every step runs in the cheapest form that is still safe for that step.

Script

When the rules are known

Ready-made or your own: hundreds of scripts already built, plus a full editor for when you need something specific. Same input, same output, every time - and it costs compute, not tokens.

Model

When the rules cannot be written down

Reading a free-text message, judging a tone, summarising a variance. Bounded turns, your own functions as its only tools, every call priced.

Person

When the consequences are real

An approval step with named approvers and a deadline. The run waits, and what they approved stays on the record.

Script

What is already built

Everything below runs today - not on a roadmap.

Authoring

Hundreds of scripts already built

Most of what you need is already in the library. When it isn't, your own developers add it - in the same platform, not a separate one.

  • A growing library of ready-made scripts: connectors, common jobs, drag-and-drop.
  • A visual graph of script, AI, condition and approval nodes - wire it up, no code.
  • Need something specific? A full script editor lives right here - and one script can import another as a library.
Assistance

Nothing starts from a blank page

Describe what you need and let a model draft it, everywhere you create something.

  • Describe a workflow; the assistant proposes a graph patch you iterate on.
  • Draft a script from a prompt - the assistant searches your existing scripts first, and writes its unit tests in the same step.
  • Chat with a model and hand it exactly the tools you choose, with the cost of every message shown.
  • Your admin approves which models a workspace may use.
Connections

Connected to what you already run

Automations start where your work happens, and read the databases you already have.

  • Schedule, on demand, webhook, or an email in your own mailbox.
  • SQL, MongoDB and Redis by name - no password in a script.
  • Results as a table, a chart, an image or rendered HTML.
Runtime

Infrastructure you actually control

Yours to own and reshape live, not a black-box SaaS runner.

  • Swap package sets or Python versions, then cut over with zero outage.
  • Your own clusters, work routed to the one with the right libraries.
  • See the month before the invoice does, and cap it yourself.
Security

Safe to hand to a security team

Most platforms ask you to trust the code you run on them. Here the boundary does the trusting.

  • Every script in a rootless container on a read-only filesystem.
  • Egress filtered on the packet; cloud metadata blocked outright.
  • Script changes come with a plain-English diff, an AI security scan and a before/after comparison - approvers read consequences, not code.
  • TOTP two-factor and OIDC single sign-on against your own provider.
Record

Everything that ran, kept

An automation you cannot audit is an automation you cannot defend.

  • Every execution recorded - inputs, outputs, logs, cost.
  • Every edit versioned, attributed and instantly reversible.
  • Everything waiting on you in one queue, counted in the navbar.
Trigger

Four ways to start a run. Not four rules.

Any workflow can start any way - on a schedule, from a webhook, by email, or by hand. These four are examples of what that looks like, not a rule about which team gets which trigger.

scheduled

IT Support

Monday 07:00, every week. It reads the release log and the incident tracker by name - no connection string anywhere in the script - has a model write the week up in plain English, and mails it out with the uptime chart attached inline, so it still renders six months later.

webhook

Product Managers

Your app POSTs a piece of user feedback to a workspace key. A condition drops the noise, a model classifies what is left, a script opens the Jira ticket - and whatever the model is unsure about waits in your approvals queue instead of being filed anyway.

email

Sales

A client mail lands in your own mailbox and starts the run. Thread and attachments go to a model that writes the one-page executive summary, which comes back as a dashboard card and a draft reply - and you approve it before a word is sent.

manual

Operations

Type a client name, press Run. The workflow splits across sanctions lists, credit data and your own CRM, merges what comes back, and returns one file carrying the sources, the score and the timestamp - readable a year on.

Palette

Everything you can drop on a canvas

Empty node

Drop in the process your team is most swamped by.

Bring it to a working session. We draw it together, and you leave with it running - not with a proposal.