NewKW5-230M: the third generation, Kiswahili firstRead
Regnant
Search/Request access
SystemsKNTTWorkflow automation

Knott. Workflow automation, with a person at the gate.

Automates business workflows on your own servers, and sends the exceptions to a person.

Talk to us about Knott How it works

What it is

A workflow has to survive a restart.

For operations and finance teams that want automation they can account for. Draw the workflow, connect the apps you use, Odoo among them, and Knott runs it from one binary on your own machines, recording the model, its confidence, its reasoning and the person who approved each decision.

A restarted process should resume its work. A failed API call should take the error route you drew. A decision that needs a person should wait for one. Knott brings those behaviours into a single binary, on a desktop or a server.

Sourceregnant-io/knott

How it works · KNTT

A decision's journey. To a person, when it should.

The route is drawn before the run. A confident decision goes through, an unsure one waits for a person, a failure takes the error route, and a restart picks up from the checkpoint.

One workflow, four runs

01 / 05COMPLETE

Confident

A webhook starts the run. The AI decision clears its threshold, the order is raised, the run completes.

Read the figure as text
  1. Confident. A webhook starts the run. The AI decision clears its threshold, the order is raised, the run completes.
  2. Unsure. Below the threshold, the decision goes to a reviewer with a deadline. The run waits for a person, then continues.
  3. Failure. The call keeps failing past its retries and timeout, so the run takes the error route that was drawn for it.
  4. Restart. The process restarts mid-run. It resumes from the last checkpoint and does not raise the order twice.
  5. Record. Whichever way it went, the record keeps the model, the reasoning and the person behind each decision.
  • Webhook trigger: Workflows start from real events: an inbound webhook, not only a manual run.
  • AI decision: A local model decides, and scores its own confidence against the threshold you set.
  • Tool call: A connector acts in the business system, here the purchase order, with its own retry and timeout.
  • Human review: Review tasks support approval, rejection, forms and deadlines. The run waits; it does not guess.
  • Retry: Each step has its own retry policy and timeout before it gives up.
  • Complete: The run reaches its end node: complete, approved, rejected or escalated.
  • Error route: When a step fails past its retries, the run takes the error route you drew, here an escalation.
  • Execution record: Every AI decision, approval and node transition is recorded with its time and justification.
  • Checkpoints: Resolved forward edges are checkpointed, so a restart can continue without repeating a side effect.

Try it · KNTT

The run is waiting. Decide, and say why.

The AI decision fell below the threshold, so the order waits for a person. The justification is required, and what you write is what the decision log keeps. The order is illustrative.
Human review · PO/2026/0418due in 04:00 h

Approve the purchase order for spare pump seals?

AI decision
Approve
Confidence
0.72 below 0.85
Reasoning
Price is 18% over the last order; stock covers nine days.
Cited by Matta
vendor V-1042, three orders this year
Write a few words first.

Waiting for a person

A process that keeps waiting on someone?

Say what it should do, which data it can touch and where it has to run. A person replies by email.

Talk to us about Knott All systems