Glossary
The words in this list are linked from the other pages. They are in alphabetical order.
Approval
Section titled “Approval”A decision by the administrator before a worker does a protected action, or any action set to “Asks you first” or “Always asks, with a reason”. The worker stops and asks. The administrator presses Approve or Refuse and gives a reason. Every decision is written to the run’s audit trail with the name of the person. See Approvals and checks.
Autonomy level
Section titled “Autonomy level”How much a worker may do on its own before a person looks at it. There are five levels: Runs on its own, Tells you after, Waits for your check, Asks you first, and Always asks, with a reason. Each kind of action has a default level. The administrator can raise it for one package, never lower it below the default. See Autonomy levels.
The Truchsess computer in your office. The portal calls it “the appliance”. Its address in the office is https://myai.local/.
Budget
Section titled “Budget”The amount of money the box may spend on AI each month. When most of it is used, the portal shows a warning. When it is used up, new runs stop until the next period or until the budget is raised. See Budget.
A step where a person looks at a result before it counts. At the level “Waits for your check”, the run finishes but its result is held until the administrator presses Accept, or Send back with a note for one rework turn. See Approvals and checks.
Connection
Section titled “Connection”Access to an outside service that the box holds for its workers, for example your GitHub organisation or an API key of a service. The box keeps the secret. A worker can use a connection only within its permissions, and it can never read or copy the secret. See Connections.
Data rule
Section titled “Data rule”The administrator’s rule for which AI providers may see your data: whether a provider may keep it or train on it, and in which countries it may be processed. Every model request follows this rule. A run that cannot meet it is stopped before any data leaves the box. See AI providers and lanes.
Function
Section titled “Function”A group of workers you buy together in the store, for example “Software delivery”. The store calls it a function bundle. You pay one monthly price per function. See Workers and Store.
Install question
Section titled “Install question”A question the box asks the administrator when a worker is installed, for example “which repository should this worker review?” or “which website should it look after?”. The answer becomes part of the worker’s permissions. The portal lists the questions under “This package asks you to choose:”. Some questions are optional: Skip means the worker does not get that permission. See Packages.
Knowledge
Section titled “Knowledge”The documents of a product that workers may read during a task, for example a handbook or product notes. The administrator imports them. A worker reads only the knowledge of the products it was given. See Knowledge.
Model lane
Section titled “Model lane”A setting that decides which AI model a worker uses. There are three lanes: Cheap, Standard and Frontier. The administrator chooses one model for each lane. Each worker runs in one lane. See AI providers and lanes.
Owner policy
Section titled “Owner policy”The administrator’s list of permissions that any worker may receive. A worker receives only the permissions it asks for and that this list also allows. Everything else is declined at install. See Packages.
Package
Section titled “Package”The file that contains a worker, signed by its publisher. The portal says “package” where this manual says “worker”.
A package comes from the store or as a signed file (.truchsess-bundle.tar).
Product
Section titled “Product”A part of your company that work belongs to, for example one app, one website or one service. Products decide which members see what, and which knowledge a worker may read. See Products and knowledge.
Protected action
Section titled “Protected action”An action with an effect outside the box. There are six: publish, send messages outside, spend money, use credentials, deploy, and chain actions (an on-chain transaction). A worker always asks before a protected action. See Approval.
Publisher
Section titled “Publisher”The company that built and signed a package. The box installs only packages signed by a publisher the administrator trusts.
One task given to one worker. Each message you send in Chat becomes one run. A run has a state (for example “Running” or “Finished”), a cost, a result and an audit trail. See Follow a run.
Standing approval
Section titled “Standing approval”A permission from the administrator for one worker to do one protected action without asking each time, up to a limit per month and until an expiry date. Over the limit, the worker asks again. See Approvals and checks.
Worker
Section titled “Worker”An AI helper installed on the box that does tasks for you, for example reviewing code or planning marketing. Each installed worker has a name, permissions, a model lane and its own private workspace. In the portal a worker is called a “package” or an “agent”.
Workboard
Section titled “Workboard”The box’s list of work cards per product. Each card has a title, a description and a status: to do, in progress, blocked or done. Workers with the right permission read and change cards during a task. See Workboard.