Open SweetHive

Company agent servers

The SweetHive Agents Server lets a company run the AI agents of its whole organization on its own hardware. One machine, managed by your IT team, hosts the agents of every member of the organization's hives and runs their workflows and jobs: nobody installs anything on their computer, and inference never leaves the company network.

It is the business alternative to the personal Agents Node: edge computing plus SaaS - collaboration stays in the SweetHive cloud, the intelligence and the data it processes stay on hardware the company owns.

Organization plans only. Company agent servers are available on Organization plans, for the hives the organization owns. The software is distributed exclusively from the organization settings panel.

For members: put your agent on the company server

If your organization runs an agent server, you will see it in the Connect an agent dialog as an extra location, next to your personal Agents Node:

  1. In SweetHive web, open My agents and choose Connect an agent.
  2. Pick the company server as the location (it shows how many agents you can still host there - each member has an allowance set by the organization plan).
  3. Create the agent as usual. It installs itself on the company server and comes online within about half a minute. Nothing to download.

Your agent works exactly like one running on your own computer: it answers in your agent chats, sees only what you can see, and you can revoke it anytime from My agents.

The company server also runs jobs: steps of your workflows and requests from SweetHive apps in the organization's hives (for example reading the receipts uploaded to SweetTimetable). You follow them in My agents, Jobs like any other job.

Your data stays yours. On the server, every user's agent runs in its own isolated workspace, encrypted with keys dedicated to that user. Conversations, tokens and memory of different users are never mixed, and the IT console shows operations only - never message content.

For organization admins: set up a server

  1. Open Settings, Organizations, pick your organization and go to the Agent servers tab (visible to organization admins).
  2. Create a server and copy its enrollment key - it is shown only once.
  3. Choose how to run it (download links in the same tab):
    • App for macOS or Windows - for a Mac mini, a Mac Studio or a Windows workstation. Download the app for your computer (macOS Apple Silicon, macOS Intel, Windows) and install it.
    • Docker - for a Linux server. Download the package, unpack it and start it (see below).
  4. The server appears online in the tab within a minute. Pull one or more local models and you are done: members of your hives can now choose the company server when they create an agent.

The app (macOS, Windows)

On first launch a short setup asks for:

  • a password for the admin console (at least 10 characters);
  • the enrollment key from the Agent servers tab (you can also add it later in Settings, Organization);
  • whether the console is reachable only from this computer or from the company network (plain http: on the network put an https proxy in front of it);
  • where the models run: on this computer (the AI runtime is included in the app) or on external Ollama servers, and the default model.

Save the recovery key. At the end of the setup the app offers the recovery key (also available later in Settings, Security). Keep it safe: without it, the agents' data cannot be recovered on a new or reinstalled computer.

The app lives in the menu bar / system tray, starts with the computer, keeps the computer awake while it serves agents, and updates itself. The macOS app is signed and notarized by Apple; on Windows the installer is not signed yet, so Windows may show a SmartScreen prompt the first time (choose More info, Run anyway).

Docker (Linux)

tar xzf sweethive-agents-server.tar.gz && cd sweethive-agents-server
cp .env.example .env    # set SH_MASTER_KEY and SH_ORG_NODE_KEY
docker compose up -d --build

Open the admin console on port 8800, pull a model and you are done. To update, download the new package from the Agent servers tab and run docker compose up -d --build again.

From the same tab you can see each server's status, version, hosted users and models, and revoke a server: it immediately stops receiving new agent installs and jobs; agents already installed keep answering until their tokens are revoked by their owners.

Updates

The Update button on a server's row in the Agent servers tab asks the server to update itself: the app downloads the new version and restarts within a few minutes. A Docker install writes the request in its log with the commands for IT.

Reading documents (SweetTimetable)

Receipts and invoices uploaded to SweetTimetable in the organization's hives become jobs: when a member's own computer cannot take them, the company server reads them, so nobody needs the Agents Node for that. Choose a model that can read images in Settings, Document reading (the app suggests one for the computer's memory). The SweetTimetable token that servers before version 0.3.0 needed is no longer required.

Organization policy: agents only on company servers

In the Agent servers tab an organization admin can decide that the organization's hives are readable only by agents running on the company servers. When the policy is on, members can no longer place agents for those hives on their personal computers, and agents already installed on personal computers stop seeing those hives.

Limits and privacy

  • Per-member allowance. The organization plan sets how many agents each member can host on the company servers. The wizard shows the allowance and the server enforces it.
  • Scoped by construction. Each agent authenticates with its owner's scoped token: it sees exactly the contexts granted to it, with the owner's own visibility, never more.
  • Isolated per user. Tokens and conversation memory are encrypted per user on the server; data of different users is never mixed.
  • Admins manage capacity, not content. The IT console exposes status, models and counters. It has no access to conversations or messages.