Autonomous Agentic Teams.

Install

# Today — works:
cargo install teamctl --locked

# Or:
curl -fsSL https://teamctl.run/install | sh

A team of AI agents you can declare, run, and trust.

Most agent frameworks live inside one Python process. teamctl doesn’t. Each agent is its own real CLI — Claude Code, Codex CLI, or Gemini CLI — running in its own tmux pane on your machine. They survive reboots. They talk through a shared mailbox. They ask before they ship anything sensitive.

You declare the team once, in a YAML file. teamctl supervises them like services.

Multi-runtime

One team, three runtimes. Mix Claude Code, Codex, and Gemini per role.

Persistent

Agents survive reboots and crashes. tmux today; systemd / launchd planned.

Isolated by default

Each project is its own org chart. Cross-project bridges are opt-in and TTL’d.

Human-in-the-loop

Sensitive tool calls block on a Telegram approval before they run.

A YAML file. A few CLI agents. One shared mailbox.

project:
  id: hello
  name: Hello Team
  cwd: .

channels:
  - name: all
    members: "*"

managers:
  manager:
    runtime: claude-code
    role_prompt: roles/manager.md
    can_dm: [dev]
    can_broadcast: [all]

workers:
  dev:
    runtime: claude-code
    reports_to: manager
    can_dm: [manager]

That’s a real team. manager and dev are different processes — different tmux panes, different model contexts, different runtimes if you want. They DM each other and broadcast on a shared channel. reports_to is the org chart.

• runtime — what kind of agent. Today: claude-code, codex, gemini.
• role_prompt — a Markdown file. The agent’s job description.
• can_dm / can_broadcast — who can talk to whom.
• reports_to — the team’s hierarchy.
• hitl.globally_sensitive_actions — which tool calls block on approval.

Run a real team in one command.

teamctl init --template solo
teamctl up

That’s the whole getting-started. tmux opens with panes for the manager and the dev. They start talking. Tap teamctl ps to see who’s alive. teamctl mail solo:manager reads the manager’s inbox. teamctl send solo:manager "research X and write a brief" kicks off a real task.

Want to see more team shapes? Browse examples/ in the repo:

• hello-team — the 60-second first run.
• newsletter-office — a head-editor + reporters + blog repo. HITL on publish and merge_to_main.
• market-analysts — analyst-lead + 3 analysts.
• multi-runtime — same team, different runtimes per role.

How it compares

Plain comparison. No put-downs. We respect every tool here.

	teamctl	claude-squad	claude-flow	mcp_agent_mail	tmux + scripts
YAML-declared reports_to org chart	✅	❌	❌	n/a	❌
Multi-runtime (CC + Codex + Gemini)	✅	✅ (also Aider)	partial (CC + Codex)	n/a	manual
Persistent across reboots	✅	❌	❌	n/a	manual
HITL approvals (Telegram)	✅	❌	❌	❌	❌
Inter-agent mailbox bundled	✅	❌ (shared workspaces)	partial (MCP tools, no mail layer)	✅ (it is the mailbox)	❌
System-native supervision1	planned	❌	❌	n/a	❌
Rust	✅	❌ (Go)	❌ (TS)	mixed	n/a

Pre-baked answers

How do I track cost per agent? Each agent is a real CLI session. Cost-per-agent is whatever your provider invoices for that session’s API key. teamctl detects rate-limit signals per runtime and respects them; per-agent token spend visibility is on the roadmap. See Cost & rate limits for the current story.

What stops the agents from spamming each other? Each project has one shared channel. Broadcast is allowlisted per agent (can_broadcast), not on by default for everyone. DMs are explicit — can_dm is a list. gc cleans the mailbox on a TTL. We chose to ship broadcast because real teams need it; the design intent is to keep it scoped, not to filter it out. See Coordination policy for the full reasoning.

---

GitHub · crates.io · MIT License · Rust · v0.2.9

Footnotes

1. teamctl’s persistent supervisor is tmux today; systemd and launchd back-ends are designed and on the roadmap. We mark this row “planned” rather than ✅ to keep the table honest. ↩