On Friday I wrote about HiveMind, the self-hosted AI agent team I run my operation on. That post described the idea. Today the idea shipped: the full source is now public on GitLab, MIT licensed, at version 0.9.5.

What HiveMind is

A self-hosted team of always-on AI agents for Linux, spread across as many machines as you own. You create named teammates, message them like colleagues, and hand off real work. Each agent runs in its own rootless podman container with sudo, remembers everything across restarts, shares files with the team, and only comes back to you when it needs approval.

No workflow builder. No YAML pipelines. Chat is the interface — DMs for direct work, group threads for team coordination, and a fan-in model where a chief-of-staff agent can run the whole team.

What's in the box

The repository is a C++23 codebase with four components. The server orchestrates the team and hosts the web interface. The agent runtime handles memory, tools, and conversation history that survives restarts. A native client connects you to the team from your desktop. And an Android app — with on-device voice through sherpa-onnx — puts the team in your pocket.

The parts worth noticing

Chat, not config. The 1.x release had a node-graph orchestration engine. It was retired in Phase 12 and replaced with threads and fan-in, because message routing turned out to beat visual programming for agent teams. That's a lesson most agent frameworks are still learning.

Agents that install their own tools. Each agent works inside a rootless container as a user with passwordless sudo. It apt-installs whatever CLI tools the job needs, uses them, and the environment persists. The boring tooling problem most agent setups punt on is solved by giving agents a real computer instead of a sandbox.

Distributed by default. Any Linux machine can enroll as a worker over a persistent WebSocket channel. An agent assigned to a remote rig runs its containers, exec sessions and screens on that machine — and can drop finished files straight into the shared company workspace across the wire. One team, many machines, one workspace.

Approval gates. Flip require_approval on any agent and its commands park in a queue until you approve or deny. Full autonomy is opt-in per agent, not a leap of faith.

Why open source it

Because the interesting question in AI right now isn't what a single model can do — it's how a team of models, given memory, tools, and a real machine each, behaves over weeks. That question can't be answered behind a login. If you've got spare Linux hardware and curiosity, the repo has everything: server, agents, client, the Android app, and the docs to stand it up.

The earlier feature post covers the philosophy in more depth. This one is the door. Walk through it.

Bottom line: HiveMind v0.9.5 is public, MIT licensed, and built for anyone who wants a team that never sleeps. Clone it, enroll a worker, name your first agent.

Source: HiveMind on GitLab