DEV Community

Bees
Bees

Posted on

Bees is public: a free desktop app that runs a team of AI agents on your computer

Bees went public today. It's a free, open source desktop app for macOS, Windows and Linux that runs a team of AI agents on your own computer, instead of one assistant in a single chat window.

What it does

You describe an outcome in plain language, in the What would you like to achieve? box on Home. Bees turns it into stages on a shared board. Agents work through those stages, and a separate reviewer agent checks each result before the work moves on.

A small example from the README: ask it to browse Hacker News and give you a table of the current stories, grouped into news, Show HN and Ask HN. Bees plans the stages, the agents do the browsing and sorting, the reviewer checks the table, and the finished result lands in that run.

You bring the AI

Bees doesn't sell you a model. Under Settings, AI connections you connect one of these:

  • Codex or Claude Code from the command line, so Bees uses a subscription you already pay for
  • a hosted provider
  • a local model on your own machine, run through a bundled llama.cpp server

You can set a model per agent or one default for everything.

What stays on your computer

Your agents, files, browser sessions, cookies and passwords stay local. Bees opens a private workspace with no sign-in, so you can try it right away. It also runs its own local memory service, so agents remember past work without a cloud memory provider.

Control over what agents can do

  • Tools per agent. Agents reach outside tools through MCP servers. You can give one agent every tool, none, or a named few.
  • Approval per stage. Mark a stage as needing your approval, and its agent can't send, post, delete or pay through a tool until you say yes. Questions, approvals and failures show up under Needs your attention.
  • File scope. An agent in a run reads and writes inside that run's own folder.

Other things it does

  • Company Brain: local search over folders and Google Drive locations you map yourself.
  • Scheduling: turn a request into a recurring process, like a weekly report.

What's rough

The Mac build is signed and notarized. The Windows and Linux builds aren't code signed yet, so Windows shows a SmartScreen warning. You need 8 GB of memory for a cloud model, or 16 GB for a local one, plus 15 GB of free disk.

Try it

Two of us built this. If you've used a multi-agent tool before, I'd like to hear what made you trust it, or what made you stop.

Top comments (0)