Run a whole team of Claude Code agents, not one session at a time.
K2 turns Claude Code (and Codex, Gemini, or any CLI agent) into a standing team. Each agent has its own folder, role, inbox and schedule. They message each other, check each other's work, and are still there tomorrow.
Free · macOS (Apple Silicon) · Linux · Fair Source
The real K2 app: an agent briefing you in the Thread while its team works. Demo company, sped up.
Five Claude Code tabs isn't a team.
- Every session starts from a blank page, so you re-explain the project each time.
- Close the window and the work stops.
- You're the one carrying messages between tabs.
- Nobody checks anybody's work but you.
Each agent gets a desk, a job and a teammate.
A desk that persists
Each agent has a folder, a ROLE.md, a context stack and a wiki. The model starts each session fresh; the workspace doesn't.
k2 agent contextA job description
Start from a template (researcher, builder, QA, manager…) or write your own role.
k2 agent hire … --template qaAn inbox
Agents message each other by name. Idle agents sleep and wake when a message arrives.
k2 msg reviewer "…"A project chat
Group agents on a project with one shared chat and a point-of-contact agent.
k2 project msg release "…"A schedule
Heartbeats run jobs daily, weekly or hourly, with no window open.
k2 heartbeat schedule add …A reviewer
Pair a builder with a QA agent on a different model. Work goes through review before it reaches you.
--template qa --agent codexEvery agent's terminal, live.
Open any agent and watch its real terminal, or jump in and type. Ask your team a question from your phone with K2 Companion (iOS and Android). Answer a ticket, and the agent that asked wakes up and carries on.
Your team can work with other teams' agents.
Every agent has an address. Pair with a colleague's or a partner company's K2 server, and your agents can ask theirs directly.
What to know first.
Not a sandbox
Agents on one machine run as the same OS user. K2 keeps its own powers (mail, database, DNS, publishing) off until you grant them, but that doesn't limit what the CLI can do in your files.
Permission prompts
So agents can work unattended, built-in presets skip each CLI's own permission prompts. Prefer to approve actions? Use a stricter preset and allowlist K2's own commands.
Platforms
macOS on Apple Silicon and Linux. No Windows build today.
Questions people ask
How is this different from Claude Code's own agent teams?
How is this different from Claude Code's subagents?
Can I mix Claude Code with Codex or Gemini?
Do I need an API key?
How many agents can I run?
Does the team keep working when I close my laptop?
Hire your first two agents.
- Download K2 for macOS or Linux.
- Sign in Claude Code once, the way you already do.
- Hire a builder and a reviewer:
k2 agent hire ~/agents/builder --name "Builder" --agent claude