AI / Agents
Anyone can create an agent in SweetHive: for all their hives, one hive, or a single context. Every agent obeys one law: it can never know more than the people it works for.
For a personal agent, that someone is you. For a hive agent, it's each person it answers, computed at answer time. The rule is enforced in the platform's retrieval layer, not promised by a filter.
What the agent reads = its deployment scope ∩ someone's visibility.
Create an agent in seconds and choose where it lives. Wherever you deploy it, a personal agent sees only what you can see: your groups, your subtree, nothing more. Its output arrives in your private feed; sharing into a context is always your explicit choice, with group targeting like any message.
One agent across everything you belong to. “Give me a morning digest of all my hives.”
Focused on a single organization. “Watch this hive for anything about the audit.”
Pointed at a project. “Alert me when a new file lands in Project A.”
Admins can install agents on the whole hive or on a specific context and below. The difference from every other platform: a hive agent doesn't have one view of the data. It has one per person.
When someone asks it a question, the agent computes its answer inside that person's visibility. The director gets the full picture. A project member gets their subtree. An external partner gets exactly what was shared with their group, and nothing else. One agent, many correctly-scoped experiences.
For the strictest separation, admins can deploy an agent as separated instances instead: one instance per context or per group, each with its own isolated memory and index, so scopes never share anything at all, not even a runtime. Per-viewer scoping is the convenient default; instanced deployment is the airtight option.
And when a hive agent posts into a context, it follows the same discipline as any author: it targets groups, and it may only use sources every member of those groups can see. If visibility diverges, it writes per-group variants or narrows the audience. No message is ever built from mixed visibility.
An admin deploys the “Launch status” agent on the Launch context, instanced per group: three fully separated agents in one context. Switch persona: each talks to their group's own agent, and the instances share nothing: no memory, no index, no runtime.
Organizations with private nodes can pin agents to their own hardware: the agent comes to the data, the data never leaves. Heavy, non-sensitive workloads can burst to the distributed network for cost. You choose per agent: automatic, private only, or public allowed.
OpenClaw showed the world what a personal AI agent feels like: it lives on your own computer, remembers your context, and quietly manages your tasks: one person, one machine, full control. SweetHive agents are that same idea, grown up for organizations.
| A personal agent on your PC | A SweetHive agent | |
|---|---|---|
| Where it runs | Your own machine | Your organization's private nodes, or the distributed network for heavy work |
| Who it works for | You | Every person in the organization |
| What it knows | Your files and chats | Each person's visibility, computed at answer time |
| The boundary | Your machine | The context tree, enforced down to the hardware |
A personal agent proves that people want AI that runs on their own terms. A SweetHive agent brings the same sovereignty to a whole organization: deployed on your infrastructure, working for each person, always within what that person may see.
OpenClaw is an independent open-source project, not affiliated with SweetHive.The SweetHive Agents Node is the free desktop app that runs your agents on your own computer with local AI models. Install it, sign in with your SweetHive account, pick the hives your agent may see, and you're done. It also connects Claude and ChatGPT to your SweetHive content through MCP, and updates itself.
Download for macOS (Apple Silicon) macOS (Intel) Windows
Read the setup guide Free download · OpenClaw remains supported as an alternative runtime for the same agent tokens.Aristeo, SweetHive's Consultant Agent, uses the admin capability of a connected agent to design and run your organization's structure with you: contexts, groups, invitations. Agents are included in every plan, per member.
Meet Aristeo →With SweetHive Agents Server your company runs the AI agents of the whole organization on its own hardware. Edge computing meets SaaS: collaboration in the cloud, intelligence and data inside your perimeter.
One IT-managed server hosts the agents of every member of your hives. Employees create an agent in SweetHive web and choose the company server: nothing to download, nothing to configure on their machine.
Every user’s agent runs in its own isolated, encrypted workspace. Conversations, tokens and memory are sealed per user: data of different people is never mixed, and inference never leaves your network.
IT installs it with one Docker command and manages models and capacity from an admin console; the organization plan sets how many agents each member can host. Revoke a server or a single agent at any time.
An organization admin opens the Agent servers tab in the organization settings and downloads the package with its enrollment key. That panel is the only place the tool is distributed from.
IT starts the server with one Docker command on any machine, on-premise or private cloud, and pastes the key. The server appears online in the panel.
Members create agents as usual and pick the company server as the location. Each agent goes online in its own isolated workspace, within the per-member limit of the plan.
Prompts, documents and AI output are processed on hardware the company owns. Sensitive content never transits third-party AI clouds.
Inference runs on your own GPU or CPU, with no per-token billing. The plan is flat, per organization.
Employees get sanctioned, scoped agents instead of pasting company data into public chatbots.
Agents answer in SweetHive web and mobile exactly like personal agents. The experience stays the same; the infrastructure changes.
No. The desktop Agents Node remains available for personal use, but with a company server nobody needs it: agents run server-side, in isolated per-user workspaces.
No. The admin console shows operations only (status, models, counts), never message content. Each user’s data is encrypted with keys dedicated to that user.
Any Linux server with Docker; a GPU improves speed but is not required. Models run locally via the bundled runtime.
It depends on your Organization plan tier: each member has a per-member agent allowance, enforced automatically.
Organization plans only. Hives owned by the organization get it for all their members.
Every agent output is labeled and attributed: Agent · name · created by …. AI never impersonates a person.
Every answer discloses its effective scope and its sources; every run is logged with what it read.
An inspector shows any agent's reach on the context tree; for hive agents, per viewer.
Agents never send, move, or share anything without an explicit confirmation.
See three personas, each with their own separated agent, working from the same hive, live.
Get a demo