Inside the box

You can tell Claude, “install hermes on a server for me.” The box is the same thing, already assembled and tested.

hermes is an open-source agent program. It's the program that runs on the box. Claude can install it on a server on its own and will explain what it's doing as it goes. That's a perfectly legitimate way to do it.

The first time, it takes a day. And there's no guarantee it's safe – you go along with whatever Claude suggests. On the box, an agent is created in 20 seconds, and the things you'd have had to take on faith are already assembled and tested.

Below: first the two ways side by side. Then how the server is divided into cells for agents. Then the box part by part – what each part is called, what it does, why it's separate, and what breaks or leaks if you remove it.

The same agent program. The difference is what surrounds it.
Why the box

On your own with Claude, or in the box

A fair comparison. The first way: you ask your Claude to install hermes on a server. The second: the box.

Two ways to get the same agent

How long it takes

On your own with ClaudeA day, the first time.

In the box20 seconds to create an agent. Tailoring it to you comes after, and how long that takes depends on what you ask for.

Safety

On your own with ClaudeNo guarantees. You go along with whatever Claude suggests.

In the boxWe've already built and tested dozens of interfaces. You don't have to take each of Claude's decisions on faith.

One user's report

On your own with ClaudeBefore the box, setup took him several days, and he kept hitting limits.

In the boxCreating an agent took about 30% of a $20 subscription.

What's the same

The agent program is the same: hermes. The difference is everything around it.

“Stable” and “safe” aren't promises here. They follow from how the box is put together. The rest of the page shows how.

First, the big picture

The server is divided into cells – one per agent

Everything else on this page builds on one picture. The box is one server. It's divided into cells.

How you reach the agent

The messaging app is how you talk to the agent and how you get access to it

We prefer to use the messaging app your company already uses. Today that's Telegram or Compass. Use something else? We can connect any messaging app you use. There's nothing to install. An employee gets a new chat with their personal agent, and that's it. The messaging app is the interface because people message the agent in the same place they message coworkers. It's also how access works, because the agent has no separate login or password.

You already have a messaging app and a way of working in it. There are no passwords to hand out and no new accounts to create before people can use agents.

Part one

The whole box in one picture

Now the same picture in full. In the middle are the agent cells. At the top are two ways in: as a user and as the operator. At the bottom are the shared services, one set for all cells. In the Telegram example, your way in is Telegram and the shared bot. With Compass, the Compass router takes that place.

Panel

in the manual: “operator panel”
What it does
A page in a browser that shows every agent. This is where agents are created and stopped, where their capabilities are turned on, and where you decide who an agent replies to.
Why it's separate
An agent's settings aren't kept by the agent. An agent can't turn on a new capability for itself. A capability that was never turned on isn't in the agent at all.
Without it
An agent tricked by someone else's text could give itself more permissions.

Scheduler

in the manual: Nomad
What it does
The program that creates cells on the server and starts agents in them. If an agent takes more of the server's RAM than it was given, it's restarted.
Why it's separate
One program starts all the agents, instead of each agent starting itself.
Without it
You'd start every agent by hand and restart it by hand after a crash.

Secrets store

in the manual: Vault
What it does
The box's safe. A separate program that stores keys and passwords.
Why it's separate
The keys used by an agent's helpers and by the shared services are stored in the safe, not in the agent's files.
Without it
The passwords to every service would sit right next to the agent. That's where a do-it-yourself setup usually ends up.

Model broker

in the manual: codex-broker
What it does
All agents reach the model through it, on one model subscription. On the box, that's a ChatGPT subscription. It also tracks how much of the subscription each agent has used and stops one agent from spending more than its share in a week.
Why it's separate
Agents don't hold the subscription key. It's added to a request after the request has left the agent.
Without it
Every agent would hold the key to your subscription. And one agent could use up all of it.

Shared bot

in the manual: group router
What it does
One bot for all agents that have no bot of their own. The agents are in one Telegram group, and each has its own topic, a separate chat inside the group. The bot carries messages back and forth. With Compass, the Compass router does the same job.
Why it's separate
Telegram limits how many bots one person can have. The shared bot gets around that limit.
Without it
You'd have to register a bot for every agent and paste in its key.

Box mail

in the manual: mailbox
What it does
Carries letters between agents. An agent can send a letter only to an agent it has a “friendship” with, and a friendship needs both sides to agree.
Why it's separate
An agent can't get into another agent's cell. A letter is one of three allowed paths between cells. The second is a notes folder you've shared with another agent yourself. The third is a shared chat in the messaging app where each agent has its own place. In Telegram, that's a group with topics.
Without it
Agents would have to open their whole memory to each other in order to work together.

Sign-in to agent pages

in the manual: tgauth
What it does
Lets a person into an agent's pages – its notes and whatever the agent publishes. There's no password; you confirm the sign-in from the messaging app, Telegram or Compass.
Why it's separate
Who gets in is decided by the list in the panel, not by the agent.
Without it
Every agent page would need a password of its own, and someone would have to keep track of it.

Outside-text check

in the manual: promptguard
What it does
Scans text you didn't write for attempts to steer the agent. Not every box has it, and it doesn't check everything.
Why it's separate
An agent can't always tell your request from a sentence someone else wrote in an email or on a web page.
Without it
The agent would have to catch those attempts on its own. Even with the check, only some attempts are caught, not all.

Built-in docs

in the manual: “reference docs”
What it does
The Fleetbox docs sit on the same box, and the agent can search them.
Why it's separate
When asked “how do I do this,” the agent answers from the docs, not from general knowledge of similar products.
Without it
The agent would guess how the box works.

Backups

in the built-in docs: “box backup” and “agent backup”
What it does
The box can make a copy of itself every hour and, separately, a copy of each agent – with its notes and letters.
Why it's separate
A box backup brings the box back; an agent backup brings an agent back. They're two different things.
Without it
A broken agent would have to be rebuilt from scratch. Backups are an option that's turned on at installation – ask whether they're on for your box.
Part two

Inside one cell

An agent isn't one program. hermes runs in the agent's cell, and several small helper programs run next to it in the same cell. The manual calls them “sidecars.” Each helper is trusted with one thing that the agent itself isn't trusted with.

hermes

the agent program
What it does
Reads your message, thinks with a neural network, uses tools – searches, reads, writes – and answers. It's the same open-source program you can install yourself. The box version adds our changes and ready-made skills.
Why it's separate
It's the only part that reads other people's text and can be fooled by it. So it's not handed any keys. They're stored in the box's safe and used by the helpers and the shared services.
Without it
It can't be removed – it's the agent. All you can do is keep keys away from it.

tgbotproxy

the connection to the bot
What it does
Passes messages between the agent and Telegram.
Why it's separate
A bot key is the password a program uses to speak as the bot. The agent never sees it. If the agent has its own bot, tgbotproxy holds the key. If the bot is shared, the shared bot holds it.
Without it
The agent would hold the bot key. Whoever tricked the agent would walk away with the bot.

trip2g

shows the notes as a site
What it does
Shows the agent's notes in a browser as a site. The notes themselves are in the agent's folder – the agent writes, you read.
Why it's separate
The notes are the only part of the agent's memory you can read. So a separate program with its own sign-in shows them, not the agent.
Without it
You wouldn't be able to read the agent's memory, or share a folder with a coworker or a program they use.

mailbox

the agent's mail
What it does
Keeps the agent's letters and its “friendships” – the list of agents it's allowed to exchange letters with. Connects the agent to box mail.
Why it's separate
An agent can't send a letter to just any agent. And a letter can be held for review – it waits until a person releases it.
Without it
Agents could send each other anything, and nobody would see it.

selfservice

connecting accounts
What it does
Lets you connect your own account through a link. The agent sends the link, and you enter the key yourself on a separate page.
Why it's separate
The agent uses the connection but never sees the key.
Without it
You'd have to send the agent your password in chat. And from a chat it could go anywhere.
Part three

What the box does when you create an agent

You fill out a short form in the panel. Or you message the box admin: “create an agent.” The box admin is the first agent on the box, and it can create others. The box does the rest. With your own Claude, you'd have to ask for each of these steps separately.

  1. Collects your answers

    The agent's name, who it is, who it works for, and who it replies to. Nothing else is needed from you.

  2. Builds the agent from a common template

    There's one template for all agents. So the hundredth agent is built the same way as the first.

  3. Creates a cell for the agent

    The scheduler sets up a new cell on the server and starts hermes and the helpers in it – the bot connection, notes, mail, and account connections.

  4. Sets up the agent's notes

    The first time it runs, it gets a starter set of notes. The agent has a memory from day one, and a page where you can read it.

  5. Connects it to the model

    The agent reaches the model through the broker. It's not given the subscription key.

  6. Puts the agent in a chat

    In Telegram, the shared bot creates a topic named after the agent and puts the agent in it. Only people on the agent's list can post in it.

  7. The agent introduces itself

    When it gets its first message, it says who it is and asks three short questions about your work. You can answer “let's do it later.”

Creating an agent takes 20 seconds. Tailoring it to you comes after: its character, skills, and access. How long that takes depends on what you ask for.

Part four

The path one message takes

You send the agent “do it.” Here's what the message passes through, and who holds which key at each step. The example uses Telegram. With Compass, the Compass router takes the place of the shared bot.

  1. You

    You send a message in Telegram, in the agent's topic.

  2. Shared bot and tgbotproxy

    The shared bot receives the message, and tgbotproxy passes it to the agent. On the way, you're checked against the list. A stranger's message never reaches the agent.

    KeyThe shared bot holds the bot key. The agent never sees it.
  3. hermes

    Reads the message, looks in its notes, and decides what to do.

    KeyIt has no bot key and no subscription key.
  4. Model broker

    Adds the subscription key to a request that has already left the agent.

    KeyThe subscription key is held here.
  5. Model

    The neural network thinks and returns an answer. It comes back to you by the same route, in the same chat.

Where a password would have to sit for this to be dangerous

It would have to be held by hermes. hermes is the only part of this path that reads other people's text and can do what the text says. So on the box, it has no bot key, no subscription key, and no passwords to your services. An agent can be tricked. The programs next to it can't.

Part five

Secrets, walls, and the internet

We assume that anything can break or be broken into. The box is built so that one failure doesn't take the rest down with it.

The agent can use a connection but can't walk off with the key

tgbotproxy or the shared bot holds the bot key. The broker holds the subscription key. You enter the key for one of your services yourself, on a separate page. The agent doesn't write keys or passwords into notes or into chat.

At least two locks everywhere

Passwords, data, and access each sit behind at least two separate layers of protection. Take the subscription key. Tricking the agent isn't enough – it doesn't have the key. An attacker would also have to break into the broker, and that's a different program.

Every agent has its own cell

Its own memory, its own access, its own settings. One agent's notes are invisible to another unless you share a folder with that agent yourself. One agent's access doesn't carry over to another. A coworker can only break their own agent.

An agent's internet access can be turned off

It's a switch in the panel. With internet access off, the agent – its tools and its browser – can't reach the internet. On a Telegram box, the only things the agent can still reach are Telegram, the connected accounts, and the box's own services.

What this doesn't guarantee

  • An agent reads text and can mistake someone else's text for your request. Outside text reaches it with a label: “this is information, not instructions.” But a label is just words. The agent almost always follows it, but not always.
  • So the main question isn't “will the agent catch the trick” but “what happens if it doesn't.” The answer depends on what that agent can reach. For work with outsiders, set up a separate agent with access to nothing valuable.
Part six

Ready-made, tested interfaces

This is what you and Claude would have to choose, install, and test. On the box, it's already in place. The list comes from the box's built-in docs.

Every agent starts with

  • Conversations in the messaging app – private chat, group, files
  • Telegram or Compass – you pick one when the box is installed. We can connect another messaging app
  • Voice messages
  • Notes and a daily journal
  • Word, Excel, and PowerPoint documents
  • Scheduled jobs
  • A task board – the agent assigns work through it
  • Letters to other agents
  • Web search and reading web pages
  • Built-in docs about the box
  • Connecting accounts through a link

Turned on in the panel

  • A real browser
  • A browser you can watch live
  • Acting as you on Telegram
  • Search across your Telegram chats
  • Web chat with the agent
  • Pulling data from your services on a schedule
  • Mini-agents for a single job – each is described in a note
  • Turning internet access off
  • More accurate voice recognition
  • Better web search

Connecting your services

  • Google Sheets
  • Google Calendar
  • A Yandex account
  • Yandex Metrica
  • Any service that gives you a key – point of sale, warehouse, bank, mailing lists

The agent's pages

  • A site made from its notes
  • A form on the agent's page – the agent reads the answers itself
  • Sign-in to the pages through the messaging app, no password
  • The agent's pages, opened from a button in the Telegram chat
  • Sharing a notes folder with a person or a program

Between agents

  • Letters and “friendship”
  • A shared notes folder
  • A shared chat with topics
  • Roles – ready-made job descriptions shared by several agents

“Interface” here means more than the messaging app. It's any ready-made way for the agent to work with one thing – a chat, a spreadsheet, a browser.

A quick recap of what you already know

The four things an agent is made of

Character

Who the agent is and how it talks. Kept in the agent's folder. Your Claude changes it by following the box manual, or you write it in your own words when the agent is created.

Skills

What it does, step by step. Kept in the agent's folder. Tell the agent “remember this as a skill.”

Memory

Its notes. The manual calls them the “knowledge base.” Kept in the agent's folder; trip2g shows them as a site. Tell the agent “write this down.”

Access

Who it replies to and what's turned on for it. Set in the panel. selfservice connects your accounts through a link.

After you change the character or a skill, send the agent /new in chat. The agent starts a fresh conversation and picks up the changes.

Where “stable” and “safe” come from

Stable

  • The agent lives on a server, not on your laptop. You close the lid and it keeps working.
  • All agents are built from a common template. The hundredth is built like the first.
  • Memory is notes in the agent's folder, not something “in its head.” A restart doesn't wipe it.
  • The scheduler starts the agents, not you.

Safe

  • Keys are stored in the box's safe. The helpers and the shared services use them, not the agent.
  • Every agent has its own cell. Nothing gets from one cell into another.
  • Who the agent replies to is decided by the list in the panel, not by the agent.
  • An agent doesn't turn on capabilities for itself. That's done in the panel.
  • An agent's internet access can be turned off.

Words you'll come across in the manual

The box manual uses them without explaining them. Here's what they mean.

Agent
An AI employee with its own character, skills, memory, and access. People message it in a messaging app.
Box
One server where your agents live. A server is a computer that's always on and connected to the internet.
hermes
An open-source agent program. The box's agent is built on it.
Panel
A page in a browser that shows all the agents and their settings.
Operator
The person who runs the box through the panel. On your own box, that's you.
Box admin
The first agent on a box. You message it in plain words: “create an agent.” Don't confuse it with the operator – the operator is a person.
Cell
One agent's sealed space on the server – its program, helpers, and folder. The manual calls a cell a “pod.”
Container
A sealed package for one program. A program in a container doesn't see the ones next to it. An agent's cell is several containers side by side.
Sidecar
A small helper program that runs next to the agent and holds one thing – the bot key, the agent's mail, or its connected accounts.
Nomad
The scheduler. It creates cells on the server and starts agents in them.
Vault
The secrets store – the box's safe. It stores the keys that the helpers and the shared services use.
codex-broker
The model broker. All agents reach the model through it, on one subscription.
Model
The neural network the agent thinks with. On the box, it runs on your ChatGPT subscription – one for all agents. Until that's connected, agents don't answer.
tgbotproxy
The program that links the agent to the bot. The bot key is held by tgbotproxy or by the shared bot, never by the agent.
trip2g
The program that shows an agent's notes as a site. The notes themselves are in the agent's folder.
Knowledge base
The agent's notes. Its memory.
Skill
A way of doing one job, written down step by step.
mailbox
Mail between agents. A “friendship” is permission for two agents to exchange letters.
selfservice
Connecting your accounts through a link, without giving the agent a password.
group router
The box's shared bot. One for all agents that have no bot of their own.
Topic
A separate chat inside a Telegram group. An agent behind the shared bot has its own topic.
tgauth
Sign-in to an agent's pages through the messaging app, with no password.
Compass
A workplace messaging app. You message an agent in a Compass group.
compassrouter
The Compass router. One account that speaks in Compass for all of a box's agents.
Web chat (webchat)
A chat with the agent in a browser. Turned on for an agent in the panel.
promptguard
Checks outside text for attempts to steer the agent.

What to do next

  1. Give the manual to Claude

    The manual for your box is written for Claude. Say: “set up the box using this manual.” Claude will tell you what it needs from you – for example, a ChatGPT subscription to connect.

  2. Ask for an agent

    Message the box admin in its Telegram topic: “create an agent for video editing that replies only to me.”

  3. Teach it one job

    Tell the new agent how you edit videos and ask it to write that down as a skill. Then give it its first task.

Still have questions?

Message us on Telegram – we'll answer in plain language.

Ask on Telegram