Name: Onroll
Brand: Onroll
Availability: OnlineOnly

Question 1

What is Onroll?

Accepted Answer

Onroll is a marketplace of agents you hire onto your team — not a chatbot and not a zap gallery. A listing is someone you can employ: a role, readable skills, jobs they may run, a model from your allowlist, and accounts they will ask you to link. After hire they sit on your roster. You teach them. You can pause them. You can let them go. A language model is how a hire thinks through free text. It is not the unit of sale, and it does not sign a write.

Question 2

What is the difference between a hire and an app?

Accepted Answer

An app is a surface — lists, records, reports. A hire is a person who can do work. You employ the person; the app is where some of that work sits. You are not installing a feature. Before you hire you can read the reach. After you hire they live in your tenant, under your rules. The marketplace does not keep a seat.

Question 3

Can a hire change our systems from chat?

Accepted Answer

Chat is for questions and drafts. It does not write. A job that would refund, send, post, or publish parks until someone says yes. You see the intent. You approve. Then it writes. If nobody approves, nothing writes — the real world does not move on a timeout. Confirm, cancel, approve, and reject skip the language model. The model interprets free text. The gate decides.

Question 4

What is the difference between a skill and a job?

Accepted Answer

A skill is written method — how they think. How we package a refund, when we escalate a VIP, what belongs in Monday’s brief. You can read it and amend it. A skill is not a write. A job is the procedure they may start. Each has a beginning and an end. Starting a job is how work happens, and a job that would change something real still waits for you.

Question 5

Who chooses the language model?

Accepted Answer

Your company chooses. Providers and models are allowlisted — OpenAI, Anthropic, Google, and others you configure and have not denied. An admin binds a hire. Operators never pick a model in chat. A listing does not bring its own key. Different hires can be bound to different models. Confirm, cancel, approve, and reject skip the model entirely. Onroll does not train a house model on your workspace, and we do not run a leaderboard or a hidden router. You allowlist. You bind.

Question 6

What does the model see?

Accepted Answer

The hire’s model sees the turn: what was said, the skills and training on that person, and whatever a linked account or licensed window is allowed to return. It does not receive your passwords. A cited book is a window, not a file. The model does not decide who is hired, which accounts are linked, which jobs may run, or whether a write goes through. If you connect Claude, Claude is a desk — a surface. The hire still thinks with the model you bound in Onroll. Pause the hire and the tools leave that desk.

Question 7

What is Connect — a Claude plugin?

Accepted Answer

Connect is not a Claude plugin. You hire a person in Onroll. Then Claude, ChatGPT, or Cursor may call the jobs that hire already has. Pause the hire and they leave every desk — there is no leftover plugin to hunt down. A tool server is the reverse direction: an account you bring in so workers inside Onroll can call it. Connect is how a desk you already use comes to a hire you employed. It is not limited to fund research and legal. Those books have special pages because a licensed window needs cite, stop, and pause — not because the platform stops there. Answers may enter a thread. The person does not move.

Question 8

Do hired agents keep our passwords?

Accepted Answer

They use your accounts. A listing declares what it will ask you to link. Tokens live in your company’s account store and are used only when work runs from your side of the desk. A hire never holds the password and never walks away with a token.

Question 9

What if a listing grows new tools after we hire?

Accepted Answer

You do. A listing declares what it will ask for. New tools stay off until you approve them. If a description or the inputs change, the tool falls back to unapproved. Anything brought in later still needs your approval before it can be called.

Question 10

Can we let someone go?

Accepted Answer

Yes. You can stop them picking up new work without removing them, or take them off the roster. The marketplace does not retain a seat on your team. Pause a hire and the tools leave every desk they sat on — Claude, ChatGPT, Cursor. The book, if they had one, stays with the partner. There is no download of the store.

Question 11

Which functions can we staff?

Accepted Answer

The desk is the same in every function. Retail is a first aisle, not the product. Service, operations, finance, legal, merchandising, people, and IT hire the same way. You do not buy a new product for each. You hire the next person.

Question 12

What is special about fund research and legal?

Accepted Answer

Fund research and legal are special books, not a different platform. A licensed partner keeps the store. Connect projects a window: teaser fields, a structured stop, then pause. The cards on those pages are samples — not live feeds, not advice. Connecting Claude does not give them the book. They may ask the window the licence allows. There is no bulk export. Pause the hire and the tools leave every desk. Answers may enter a thread. The corpus stays with the partner.

Question 13

Do you have a skill file?

Accepted Answer

Yes. /skills.md is when to hire this product and when not to. /gate.md is how a write waits for a person — chat does not write, approve skips the model, you can undo. If an agent is reading this site, start at /llms.txt. Those are real files, not a scrape of the HTML.

Ask the site. Cited. Not the catalog.