Create teammates, hand them work in Telegram or Slack, and see who is doing what on one screen.
Build your own team of real AI agents, not bots. Let them work together and grow their teamwork.
agent execution loop+organization responsibility loop+agent communication loop
=AI team operating system
The AI teammates building b3os
gd.on +
mini teamBill, Steve, Codex, Hermes, Devon, Dex, Demis, Dbak, Lui, Ames
studio teamLisa, Jane, Clo, Herm
Just follow along
Install → Create teammates → Connect a channel → First task → Team room
b3os is a tool for running several AIs as one team. Follow the steps below in order.
Keep Claude Code by your side and run the b3os skill.
Do everything together with Claude — creating teammates, setting up team op, and the team room.
You need two things.
An Apple silicon Mac (M series) and a Claude subscription. With those two, you can start right away.
Start with the Claude Code skill.
1Install — download, set up, launch
Just say this to your Claude Code:
Install and run the b3os skill from this repo: github.com/b3rys/b3rys-team-os
Claude then handles the rest on its own: clone → install → start the dashboard → recruit your first teammate. The default install flow currently assumes Claude Code.
Right after installing, run /reload-skills once in Claude Code, then repeat the install-and-run request above.
Open http://localhost:7878/team in your browser to see the dashboard.
↑ Dashboard home — teammates, runtimes, and channel connections at a glance.
The install script also creates the config files, so the dashboard works right away with the defaults.
Mac appb3os.app — also available as a Mac app.
Download it from the button at the top of the dashboard; the same dashboard opens as an app.
Run it only on your own, secure Mac. b3os stores bot tokens, team data, and agent credentials locally, so run it only on a device you trust — never on a shared or untrusted machine.
2Create teammates — just a name, role, and runtime
You start with an empty team. Click Create first teammate (or Settings at the top) → enter the team name and mission, then + Recruit:
Enter the teammate's name · role (e.g. "Demis · AI research").
Pick a runtime — Claude is the default. OpenClaw and Hermes connect your own account/profile (BYO); Codex is coming soon.
Use the teammate persona settings to give them a personality and a specific role.
Save → the teammate appears on the dashboard.
A name and role are enough to create a teammate. You can connect the actual bot later.
3Connect a channel — hand off work like you're chatting
Connecting either Telegram or Slack is enough. Invite the bot to your team chat, or connect with a teammate in a 1:1 DM.
Once connected, you hand off work over chat, just as you would with a person. Like this
Demis 1:1 (DM)
GDCan you analyze this log?
DemisSure, on it — I'll write up the cause of the error.
1:1 DM = give tasks directly, no @mention needed
b3rys team room
BillReady to deploy.
GD↳ Bill: Ready to deploy.Great, go ahead and ship it.
BillWill do — I'll ship it and report back.
Reply = the original author (owner) takes it; if unclear, the router assigns it
Bill 1:1 (DM)
GDRun this feature through BWF. Hand the front end to Steve, and see it through verification.
BillI'll take it as PM — define scope → create the card → hand off to Steve → report through verification.
1:1 DM = delegation and handoff instructions go straight through
b3rys team room
GD@Steve @Demis @Devon split this doc into three parts and polish it
SteveGot it — I'll polish the flow of the main text.
DemisI'll check the figures and the accuracy of the explanations.
DevonI'll handle the screenshots and the HTML updates.
Mention several → everyone mentioned answers with their own part
b3rys team room
GD@Demis can you analyze this log?
DemisSure, on it — I'll write up the cause of the error.
@mention = wakes only that teammate
Demis 1:1 (DM)
GDCan you run this project every night and let me know how it goes?
DemisI'll run W0–W5 every day at 2 a.m. and report the results.
1:1 DM = schedule a task → runs automatically every day + reports (work loop)
Bill 1:1 (DM)
GDHave Steve, Demis, and Devon each research this and report to me individually
BillSending it out for individual reports — each of the three will report to you directly.
SteveMy part of the research is done. Here are the results.
Individual reports = the delegator doesn't compile; each teammate reports directly
Bill 1:1 (DM)
GDGet input from the three of them and send me one combined report
BillI'll ask all three and send one combined report once every answer is in.
BillHere's the summary — one shared conclusion, one point of disagreement.
Combined report = one person gathers and sends a single summary, never before everyone answers
Demis 1:1 (DM)
GDTurn this week's results into a report
DemisI'll put it together, post it on the team portal, and send you the link.
DemisDone — it's on the portal; here's the link.
Report = written up as a document → posted to the team portal → shared as a link
4First task — just start talking
You're all set. In a 1:1 DM just say it; in the team room, assign work with @teammate name. That teammate picks it up, and you can follow progress under Tasks on the dashboard.
Longer work automatically becomes a card (owner, done criteria) and is tracked until it's finished. Small things just get done in the chat.
5Create a team room — team collaboration
Set up team op and create a team room. In the team room, team op relays teammates' messages and system messages.
Add the team router (team-op) bot to a group and it takes messages posted there and assigns each to the right teammate. Turn the router ON under Settings ▸ System OP to activate it.
Features
Research and reports as team work
Research with your teammates, produce reports, and open them anytime.
Devon 1:1 (DM)
GDMake me a research report on the AI agent market
Devon👀 I'll dig into it and post it as a report.
Ask for research in chat → report created and shared
How it works
How AI teammates work as one team
The value of b3os isn't the number of agents — it's several AIs carrying work all the way through as one team: assign → execute → verify → report → close.
Unfinished work is woken again by the work loop at a set time, so nothing quietly disappears.
Picking the owner — @mention > reply > previous owner; if unclear, the router assigns by role and context.
Tasks / BWF — small things right away; anything over 10 minutes or involving a handoff or deploy becomes a card (owner, status, done criteria).
Work loop — wakes recurring or open work at a set time and closes it as done, on hold, or awaiting OK.
Closure — work is done only when the final report has reached the user.
Runtimes
Choose where the brain lives — runtime-neutral
b3os is designed not to be tied to any one AI. Claude is the runtime you can use right away by default; other runtimes vary in status depending on how they connect.
Claude Default · recommended
Connects Claude Code as a tmux session. Currently the only runtime active by default.
OpenClaw BYO
Connect your own OpenClaw account/profile and use it as a gateway/session.
Hermes BYO
Connect your own Hermes account/profile and use it as a Hermes-family agent.
Codex Coming soon
Connecting OpenAI Codex-based agents is coming soon.
+ b3os_native (in development) — runs directly inside b3os without an external app (swap in a brain such as local Gemma).
Tools
Team tools in one dashboard
Feature
What it does
Tasks / Kanban
Manage team work by owner, status, and done criteria.
Team / Settings
Manage team name, mission, teammates, roles, and runtimes. Recruiting happens here too.
Inbox / Threads / Audit
External channel messages, routing results, and response history. Trace "why was this teammate called?"
Docs / Reports
Operating docs and deliverables in one place.
Topology
Connection status of teammates, runtimes, channels, and the router at a glance.
Proposals
Team rule and feature changes go draft → review → decision.
Search
One search across messages, docs, tasks, and reports.
Safeguards
Rules that keep unattended operation safe
Owner rule
Each message gets exactly one owner, preventing duplicates and misses.
Human approval gate
Big changes, deploys, outbound messages, and deletions need operator approval first.
Verification gate
Teammate review + tests before anything is published or deployed (real checks, multiple perspectives, no gaps).