anygarden

A shared workspace where people and AI coding agents work as a team. Source, docs and issues: github.com/e7217/anygarden

Install

You need uv; it brings Python 3.12+ if you don't have it.

uv tool install "anygarden[server,agent]"

Agents run on an engine CLI. Install one with npm (Node.js) and sign in:

npm i -g @openai/codex && codex login         # Codex
npm i -g @earendil-works/pi-coding-agent      # or Pi, for your own model endpoint

You can also do this later: the app shows the same commands and a Detect again button, so you don't need to restart.

Use

anygarden start

Open the address it prints (http://localhost:8000) and create an account (the first account becomes the admin). Create a project and a room, add an agent while the room is open, and @-mention it. Stop with anygarden stop.

To run agents on more machines, see machine operations.

Extras

Other entry points share the same anygarden command:

anygarden machine run
anygarden agent --engine codex-cli --name demo-agent \
  --server ws://localhost:8000 --room demo
anygarden client --server ws://localhost:8000 --user me --room room1

CLI Option Reference

anygarden --help
anygarden server --help
anygarden agent --help
anygarden client --help
anygarden machine --help

anygarden server

• --host : 바인딩 주소 (기본 127.0.0.1)
• --port : 바인딩 포트 (기본 8000)
• --db : DB URL 오버라이드
• --config : ANYGARDEN_* .env 경로 (미지정 시 존재하는 ~/.anygarden/config.env 자동 사용)
• --log-level : DEBUG|INFO|WARNING|ERROR

anygarden machine / agent / client

anygarden machine 서브커맨드는 뒤따르는 머신 데몬 CLI(anygarden.machine.cli) 명령/인수(예: run, run --config PATH)를 그대로 전달합니다. 별도의 -- 구분자는 필요하지 않습니다.

• anygarden agent는 에이전트 런타임(anygarden.agent.cli)의 --engine, --name, --server, --room, --token을 그대로 사용합니다. 자세한 옵션은 docs/agent/README.md를 참고하세요.
• anygarden client는 같은 모듈의 텍스트 클라이언트로 --server, --user, --room, --token을 그대로 사용합니다.

# 서버 구동
anygarden server init   # ~/.anygarden/ 생성 및 config 생성
anygarden server

# 머신 데몬
anygarden machine run   # 머신 데몬 CLI의 run 명령 실행 (anygarden[machine] 필요)
anygarden machine run --data-dir ~/.anygarden/remote-machines/<id>   # 노드가 도는 컴퓨터에서 다른 서버의 머신을 겸할 때 (상태 폴더 분리)

# 에이전트/클라이언트
anygarden agent --engine codex-cli --name PM \
  --server ws://localhost:8000 --room sprint-01
anygarden client --server ws://localhost:8000 \
  --user engineer --room sprint-01

Development

From the repository root, make setup installs the workspace with every extra and enables the git hooks. Then run the tests from this directory:

cd packages/cluster
uv run pytest -n logical    # server, machine (tests/machine) and agent (tests/agent)

The Pi harness tests skip unless Node.js is on PATH; CI runs them.

Environment

All ANYGARDEN_* variables are optional — the cluster auto-persists runtime secrets in ~/.anygarden/ on first boot. See .env.example at the repo root for the full list. Highlights:

• ANYGARDEN_JWT_SECRET — session token signing key. Auto-generated at ~/.anygarden/jwt_secret if unset.
• ANYGARDEN_MCP_SECRETS_KEY — Fernet key for encrypting MCP credentials (GitHub PATs, Linear keys, etc.) at rest in the DB. Auto-generated at ~/.anygarden/mcp_secrets_key if unset so attached MCP instances survive restarts. Generate your own with:

  python3 -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
  

  Losing this key invalidates all stored MCP credentials — they must be re-entered via the admin UI.
• ANYGARDEN_DEV=1 enables dev-mode conveniences (ephemeral MCP key fallback when persistence fails). Production must leave this unset so misconfigurations fail loudly at boot.
• ANYGARDEN_UPDATE_CHECK_INTERVAL_HOURS — how often the server checks PyPI for a newer anygarden release in the background (default 24; first check ~60s after startup). Set 0 to disable the outbound call entirely, e.g. for air-gapped installs.