Dispatch

FRONT AND CENTRE

Dispatch — the assistant that removes the writing-down

DISPATCH · WORK ASSISTANT

Send one line.
It lands on the right task.

There is no shortage of task managers. What is left over is choosing which item to write it under, and because that judgment falls to a person, nobody writes anything. Dispatch took the choosing away.

Team messenger · 4:12 pm

Sending the brochure draft by tomorrow afternoon

Product brochure revision

Due Sep 15 (Tue)In progress

Next action

Send the brochure draft

Progress log

Sending the brochure draft by tomorrow afternoon

No task was picked, no date was counted. The assistant finds the matching one among the open tasks — it asks back only when two or more have similar names.

A folder linked to a task · commit pushed

Nothing was reported

Kanban board improvements

Progress log that evening · recorded automatically

[commit] card drag-and-drop · 6 tests added

Link a task to the folder you worked in with Claude Code or Codex, and the commits made there become the progress log that evening. It does not care which tool made them.

The task list lives as md files, and the issue tracker keeps receiving your weekly hours as before.

Dates, order and capacity are worked out in code, so the answer is the same every time. The model is used once, only to read the sentence, and one task is one md file — it still opens after you delete this tool.

Dispatch in depth →

A kanban board split into seven columns, urgent ones on the left, with a red outline on the cards that need work today

95.0%accuracy spotting work in mail

3.3sphone tap to board update

$30/mowhat the model costs

Built the same way, for a different problem

The problem differs; the rules do not — no model where there is nothing to judge, and a person approves anything that leaves the building.

MYCELIUM · INTERNAL KNOWLEDGE BASE

Connects scattered company material and answers in one line

Company material is scattered across work mail, an issue tracker, a document tool, code review, file servers and more, each with a search box that only sees itself, and commercial AI search has none of this data and stays generic. Mycelium pulls that material into one index and builds a graph of how people, projects, issues and documents connect, so a question no single tool could answer comes back as one reply with its sources.

Where you use it: Work chat and the terminal — no new screen to open
Done for you: Indexed every night before dawn, taking only what is new · routed to search by meaning, by relationship, by counting and filtering, or by exact token
What you can see: Collected under your own account — your own material is connected in too, and only when you search
Deliberately out: Inventing an answer — when nothing is found it says not confirmed within the search scope
In technical terms: Hybrid retrieval over a knowledge graph — vector, full-text, SQL aggregation and graph traversal, chosen by query routing

Mycelium in depth →

"Who fixed PROJ-1234?"

By exact token (FTS5) → by relationship (KG)

Who fixed it, what they raised, and the sources, in one reply

"What is the travel expense policy?"

By meaning (vector)

Points to the clause in the company policy held in the groupware

98.1%accuracy on a regression set grown from real questions

4–12squestion to answer at 50k documents

eightsource types connected so far — more can be added

Where Dispatch and Mycelium meet

One turns the work you are doing into a record; the other pulls answers back out of the records that piled up. They are the front and the back of the same flow.

FRONT The record walks in from where the work happens You drop one line. Dates, ordering and progress are filled in by arithmetic and by your own commits.
BACK One line asks across scattered company material Eight windows and a manual merge collapse into one question.
ONCE LINKED "What happened back then?" becomes a one-line question too Mycelium already reads mail, issues and documents — the records that work leaves behind. Only the progress log Dispatch keeps separately is still outside the index, and adding it is the next piece of work.

SIDE BY SIDE

Dispatch and Mycelium handle different time

Both remove work a person was doing by hand while moving between tools. What tells them apart is which direction in time they face.

Dispatch and Mycelium on the same axes · as of 2026-09-03
Axis	Dispatch	Mycelium
Faces	The work happening now	Everything already piled up
Manual step removed	Writing up what you did	Opening and merging scattered sources
What you do	Drop one line	Ask one line
What comes back	What to start today	One answer, with its sources
Surface	A single board — same view on phone and desktop	No screen — one command and work chat
Where the model is used	Understanding a sentence, judgement — $30 a month	Meaning search, relation extraction — about $15 a month ($0.5 a day)
Where it runs	Cloud, on-premises or a local server — the store stays where you install it	Cloud, on-premises or a local server — the store stays where you install it

WHAT BOTH BUILDS TAUGHT US

Four rules that transfer to any team

Two different problems, and the rules left over are the same ones. They are about method, not about these tools, so they travel.

Move the moment of asking

The same question, asked at the moment someone can actually answer it, gets an answer. In Dispatch a field that had stayed empty for 15 days filled in a single day once the moment of asking moved.
No judgement, no model

Working dates back from a deadline and counting rows are jobs for code. What these tools hand to a model is understanding a sentence and making a call — which is why they cost $30 and $15 a month.
When unsure, ask instead of dropping

A wrong suggestion costs one line of annoyance. A wrong drop leaves no trace at all, so the miss itself disappears.
Any action that reaches the outside is approved by a person

Creating a task, posting hours, writing to a calendar — where undoing is a human job, automation stops at the proposal.

MEASURED

Every figure here came from
an installation in daily use

Nothing was produced for a demo. Each figure carries what was measured and when.

95.0%Dispatch — accuracy at recognising work in mail, graded by hand over 140 threads2026-09-03
3.3sDispatch — from a tap on the phone to the board showing it2026-09-03
$30/moDispatch — what the model costs2026-09-03
98.1%Mycelium — measured on a 417-question regression set grown from real questions2026-05-20
4–12sMycelium — one-line question at ~50,000 documents; slowest seen 45s2026-09-03
$15/moMycelium — what the model costs, about $0.5 a day2026-09-03

When material grows past a measurement date, the old figure stays until it is measured again. Scale differs from site to site, so per-installation counts are never used as figures.

The fastest way in is to click something

The demo is drawn by the real code. Only the tasks are made up — the cards actually move.

Open the demo Dispatch in detail Mycelium in detail 한국어

Contact

Ask us anything

Adoption, a demo, how it was built — whatever is useful to you.

Reply: Sent to the e-mail address you leave
What you write: Used only to reply · never passed on
See it first: Open the demo →

Do the work.The record writes itself.