Self-hosted · any model · your machine

Everything an AI assistant needs around it.

The clever part is already solved — you can buy a good model by the hour. What is missing is everything around it: what it is allowed to do, who signs off before it does anything serious, what it costs, and what it actually did. That is what this is.

Install it Documentation

curl -fsSL https://heddled.com/install.sh | sh

irm https://heddled.com/install.ps1 | iex

A conversation written out in plain words: priya@acme.example
asked whether invoice F-2231 was paid; the assistant used lookup
invoice and got back unpaid, 249 euros, Acme BV, due 14 August;
it replied with those details. — Every conversation keeps a record: what was asked, what it checked, what it found, what it said back.

A conversation paused: the assistant wants to issue a refund of
249 euros for a duplicate payment, with Approve and Refuse
buttons waiting. — It found a double payment and stopped. Nothing happens until somebody approves it — here, or from the link it sent.

Setting up an assistant: a box for how it should behave, and the
things it is allowed to do shown as removable labels. — Setting one up: how it should behave in plain language, and what it is allowed to do.

The same conversation as a numbered list of steps, with the exact
result the invoice lookup returned shown beside it. — When you need the detail, every step is there — including exactly what each action returned.

A list of conversations, each showing which assistant handled it,
what started it, and how it ended. — Everything that has happened, in one place — however it started.

A terminal showing git diff on an agent file: two lines added
that gate the create_ticket action behind approval. Nothing
else in the file has changed — comments and formatting are
exactly as they were written. — Somebody ticked a box in the browser. This is the whole diff — reviewable, revertable, and yours.

It stops and asks. Mark anything you would rather approve and the run waits for a person — enforced by the platform, not requested of the model. It carries on from exactly where it stopped.
Limits that hold. Cap what it spends in a day, how often an action may run, and keep card and account numbers out of the record.
Everything it did, written down. What was asked, what it looked up, what came back, what it decided — readable months later.
It starts itself. On a schedule, when a file lands, or when email arrives in a mailbox you point it at.
It reaches your systems. Look a record up, call an API, send a message, work through a folder of files — most of it without writing code.
It stays yours. One container on your own hardware, any model provider or one you run, and every setting a plain file you can read.

How it all works, in the documentation · a page for whoever signs things off · assistants that call each other · tests made by the person who spotted the problem