Dispatch

A Kanban board where every card is a task for an AI agent.

Try the live demoNo signup · sample data

Status: In development
Category: AI agents
Platform: Web app
Role: Product, design & engineering
Year: 2026

DDispatch / Growth board

Running2

Summarise competitor pricing page

fetch_url · acme.io/pricing

Competitor00:24

Extract contacts from Q3-report.pdf

read_file · page 4 of 12

Extraction00:11

Review1

Follow-up email for Atlas Foods

Awaiting approval

Email draft€0.03

AI agents · Web app · 2026

Drag a card to Running and an agent picks it up, works in the open, produces an artifact, and waits for a human before anything irreversible happens.

[01]

Handing work to AI today usually means a chat window and blind trust. You can't see what the agent did or what it cost, and you can't stop it before it emails a client. Teams want to hand off real work, like research, extraction and drafting, without losing control of it.

[02]

Dispatch makes agents feel like teammates on a board you already understand. Work is a card, running is a column, and every step streams into the inspector. Safe tools run straight away. Anything with side effects pauses the run and asks a human first, with the exact payload on screen.

The 20-second demo

The path a first-time visitor takes through the live demo, from the first click to the moment it clicks.

Follow along in the demo

[01]
Open the demo workspace. No signup.
[02]
Three cards wait in Backlog: summarise a competitor's pricing page, extract every contact from a PDF, draft a follow-up email.
[03]
Drag all three to Running. Three agents start in parallel and their traces stream live.
[04]
Within a minute each card lands in Review with an artifact. The email card asks for approval before sending.
[05]
Approve one, edit one, re-run one with a comment. Done.

What it does

We can build agent systems that do real work safely: orchestration, tool use, streaming and human-in-the-loop.

[01]

Board-native runs

Drop a card in Running and an agent starts. Queues, concurrency limits and retries are handled by the board, not by you.

[02]

Live trace

Plans, tool calls, results and costs stream in real time. Nothing happens off screen, and the cost counter ticks as it goes.

[03]

Approval gates

Side-effecting tools pause the run. Approve, edit then approve, or reject with a reason. A, E and R on the keyboard.

[04]

Artifacts that ship

Documents, tables, JSON and email drafts render cleanly, cite their sources, and export to Markdown, CSV or DOCX.

[05]

Comment to re-run

Mark a comment “Send to agent” and the card re-runs with your note as an extra instruction. Every run stays in history.

[06]

Five task types

Research brief, competitor summary, data extraction, email draft and structured extraction, each with its own schema and cost estimate.

How it behaves

Nothing hidden. Every agent action is visible in the trace, and every cost is shown before and after the run.
Nothing irreversible without a human. Approval gates are a feature, not friction, so they're built to be fast.
The board is the truth. A card always reflects its real state. The inspector adds depth but never contradicts the board.

V1 targets