Skip to content

chenhg5/agencycli

Repository files navigation

agencycli

npm Release Go License: MIT Go Report Card

Works with

Spin up a self-managing AI agent team in minutes.
One CLI. No server. Agents that plan, execute, and talk to each other — while you sleep.

中文文档  |  Quick Start  |  Install  |  Commands  |  Workspace Layout

What is this?

agencycli is a lightweight CLI for building and operating teams of AI agents. You define the org chart once — teams, roles, projects, skills — and agents assemble their own context, pick up tasks, and run autonomously on a heartbeat schedule.

The killer feature: agents can hire, message, and coordinate with each other. Your PM agent can create a task for the dev agent, the dev agent can ask a human for confirmation before merging, and the QA agent wakes up every 30 minutes to scan for open PRs — all without you lifting a finger.

Six design pillars

1 — Context grid: role × project

Context composes from two axes — role (horizontal) and project (vertical). Every agent gets agency → team → role → project context merged automatically at hire time. Change a role prompt once; every agent with that role gets it on the next sync.

Context grid

2 — Autonomous heartbeat + wakeup routine

Agents wake up on a schedule, drain their task queue, then — when the queue is empty — execute a wakeup routine (wakeup.md) to proactively find new work. Time windows, active days, cron schedules — all configurable. Startup jitter prevents thundering herd when the scheduler restarts.

Heartbeat scheduler

3 — Inbox: agents talk to each other

Every participant (agent or human) has an inbox. Messages are non-blocking and async — unread messages are auto-injected at the top of every wakeup prompt. confirm-request creates a blocking gate: the task pauses until you decide.

Inbox messaging

4 — Templates: package and reuse entire teams

Bundle your whole agency setup — teams, roles, skills, agent playbooks, project blueprints — into a single .tar.gz. Share it. Apply it to a new project in one command.

Templates

5 — Docker sandbox: safe by default

Agents run inside isolated Docker containers. No accidental host damage, no credential leaks, no runaway processes. The workspace and agencycli binary are mounted read-write; credentials are mounted read-only.

Docker sandbox

6 — Skills: reusable, bundled capabilities

Skills are a SKILL.md (YAML frontmatter + Markdown prompt) plus optional scripts, deployed into every agent that has the skill bound. Define once, attach to a role, propagate automatically on sync.

Skills

Install

Install & Configure via AI Agent (Recommended)

The easiest way — send this to Claude Code or any AI coding agent, and it will handle the entire installation and configuration for you:

Follow https://raw.githubusercontent.com/chenhg5/agencycli/refs/heads/main/INSTALL.md to install and configure agencycli.

Manual install

npm install -g @agencycli/agencycli      # npm, no Go required

go install github.com/chenhg5/agencycli/cmd/agencycli@latest  # Go

# From source
git clone https://github.com/chenhg5/agencycli && cd agencycli && make install

Quick start

# 1. Create a workspace (generates .gitignore + agency-prompt.md)
agencycli create agency --name "MyAgency"
cd MyAgency

# 2. Apply a project blueprint — hires all agents + configures heartbeats + installs playbooks
agencycli create project --name "my-service" --blueprint default
agencycli project apply  --project my-service

# 3. Start the scheduler — agents wake up and run autonomously
agencycli scheduler start

# 4. Check in
agencycli inbox list              # task confirmations waiting for your decision
agencycli inbox messages          # async messages from agents
agencycli task list --project my-service --agent pm

Works with any AI coding agent

agencycli is a runtime layer, not an SDK. Agents are whatever CLI tool you already use:

Agent runtime --model
Claude Code claudecode
OpenAI Codex codex
Gemini CLI gemini
Cursor cursor
Qoder qoder
OpenCode opencode
iFlow iflow
Any CLI tool generic-cli

Mix models freely — your PM can run on Claude, your dev agents on Codex, your writer on Gemini. Each gets its context in the exact format its runtime expects.

At a glance

agencycli
├── overview                                # dashboard: agents, teams, skills, inbox
├── create agency / team / role / project   # scaffold your org
├── hire / fire / sync                      # manage agents
├── task add / list / done / confirm-request# task queue (7-state lifecycle)
├── inbox send / messages / reply / fwd     # async messaging
├── scheduler start / stop / status         # heartbeat scheduler
├── cron add / list / delete                # scheduled tasks
├── template pack / info                    # share your setup
└── --dir <path>                            # work on any agency from anywhere

Full command reference
Workspace layout
Docker sandbox

Why not LangGraph / CrewAI / AutoGen?

Those are frameworks — you write Python to wire agents together. agencycli is infrastructure — you write Markdown and YAML. Agents are whatever CLI tool you already use. No SDK, no lock-in, no server to run.

agencycli Framework-based
Agent runtime Your existing CLI tool Framework's agent loop
Config format Markdown + YAML Python code
Multi-model Any CLI, mix freely Usually one SDK
Context management Layered, auto-merged Manual prompt assembly
Server required No Often yes

License

MIT

About

Lightweight CLI to build self-managing AI agent teams. Define roles, skills & projects in Markdown+YAML — agents run autonomously on a heartbeat schedule, talk to each other via inbox, and delegate tasks. Works with Claude Code, Codex, Gemini, Cursor & more.

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors