Solutions / Local AI agents

Local AI agents, built into your team's work.

With SweetHive, every team member gets personal AI agents that can run on local models: on their own computer with the free Agents Node app (macOS and Windows), or on a server your organization controls with Agents Server (Docker, Mac or Windows; Organization plans). Each agent reads only what its owner can see. The workspace itself stays in SweetHive's cloud, hosted in Ireland.

What "local" means here

The model is local. The workspace is not.

Two different things get called "local AI". In SweetHive the AI model runs on hardware you control, so prompts and the content the agent reads are processed there. Messages, files and permissions live in the SweetHive workspace, hosted on AWS in Ireland (eu-west-1). The agent fetches only what it needs from the workspace, within its owner's visibility.

Two ways to run it

On each computer, or on one server for everyone.

Both run the same agents with the same boundaries. The difference is who installs and maintains the machine.

Agents NodeA free desktop app for macOS (Apple Silicon and Intel) and Windows. Each person installs it and runs their own agents on their own computer, with local models included in the app. Available on every plan, Free included.
Agents ServerA server your organization runs, hosting the agents of all its members: Docker on Linux, or a desktop edition for Mac and Windows. Nobody installs anything on their own computer. Included in Organization plans only.
Company policyWith Agents Server, an administrator can require that agents run only on the company's servers. Agents installed on personal computers then stop having access.
What an agent can read

The owner's view, never more.

Personal, not shared

An agent belongs to one person. It works in the scope that person chooses (all their hives, one hive or one context) and never sees more than they can see at that moment.

Messages and attachments

It searches the messages of its contexts by keyword and reads the attachments of the messages it can see: PDFs, scans, images and text files up to 20 MB. Office files and links are not read today.

It asks before it posts

The agent prepares a post or an invitation and waits for your confirmation. Whatever it publishes goes out under its owner's name.

Limits and prerequisites

What to know before you start.

HardwareA GPU is not required. Plan for at least 6 GB of memory for the smallest model; more memory lets you run larger and better ones. A GPU speeds things up.
Operating systemsAgents Node runs on macOS and Windows, not on Linux. Agents Server runs on Linux with Docker, or on a Mac or a Windows machine.
When data leaves your machineWith local models, prompts and content are processed on your hardware. Agents Node also lets a person use their own Claude account instead of a local model: in that case the prompts go to Anthropic. Agents Server has no such option.
QualityLocal open models are less capable than the largest cloud models, and speed depends on your hardware. Search is by keyword, not by meaning.
Not fully on-premisesOnly the AI runs on your hardware. The workspace is a cloud service.
Example scenario

A twelve-person studio, one Mac mini.

This is an illustration, not a customer story. A design studio on an Organization plan puts Agents Server on a Mac mini in the office. Each of the twelve people gets their own agent there, with nothing to install. The project lead's agent can summarize every client context; a junior designer's agent only sees the two projects they work on. A freelancer invited as a guest to one project sees that project only. The administrator turns on the "company servers only" policy, so no agent runs on personal laptops.

FAQ
Do I need a GPU?

No. The models in Agents Node run on a normal recent computer with enough memory (at least 6 GB for the smallest one). A GPU makes answers faster, and it is needed only for the optional video rendering workers.

Are my documents sent to external AI services?

Not when you use local models: the agent reads the content it is allowed to see and processes it on your hardware. If a person connects their own Claude account in Agents Node, their prompts are processed by Anthropic.

Which models are supported?

Agents Node includes a catalog of open models and suggests the ones that fit your hardware. The catalog is updated over time, so we do not list model names here: the app shows the current list.

Does it run on Linux?

Agents Server does, with Docker. The Agents Node desktop app is available for macOS and Windows only.

How are agents updated?

Agents Node updates itself, and waits for running jobs to finish before restarting.

Who can use Agents Server?

Organizations on an Organization plan. The hives owned by the organization get it for all their members.

Read more: Agents Node · How agents work · Private AI workspace · Organization plans · Agents Node documentation · Agents Server documentation

Start with one agent on your own computer.

Create a free account, download Agents Node and connect your first agent. Move to a company server when the whole team is ready.

Try SweetHive free for 14 days