CrewLine

YAML-configured AI dev team: define agents and tasks in config, point at any project, get code + review + tests.

Русская версия ниже / Russian version below

Why CrewLine?

• No Python needed for config — new agent = a few lines in YAML. Core logic lives in main.py.
• Any project — point --project at any codebase and go.
• Two modes — agents build the plan, or you provide your own.
• Sandboxed — agents can only touch files inside the project directory.
• Budget-friendly — expensive models for thinking, cheap models for mechanical tasks.
• Dry-run — preview the execution plan without spending a cent.

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Create .env with your API key
cp .env.example .env
# Add your ANTHROPIC_API_KEY to .env

# 3a. Run interactively (no arguments needed — wizard will ask everything)
python main.py

# 3b. Or pass everything via flags
python main.py --project ./projects/example_project --goal "Add JWT authentication"

# 4. See available agents and tasks
python main.py --list-agents
python main.py --list-tasks

Architecture

CrewLine/
├── main.py              # Single entry point (CLI)
├── AGENTS.md            # Instructions for AI assistants working on CrewLine itself
├── config/
│   ├── agents.yaml      # Agent roles (edit here)
│   └── tasks.yaml       # Workflow stages (edit here)
├── tools/
│   └── file_tools.py    # Sandboxed file tools for agents
├── projects/            # Your projects (each in its own folder)
│   └── example_project/
└── .env                 # Your API keys (don't commit!)

Tip: Add AGENTS.md (or CLAUDE.md) to your target project — agents read it automatically as extra context. Describe architecture decisions, conventions, things agents must know.

Two Modes of Operation

Mode 1 — Agents build the plan (default)

The architect analyzes the project, creates a plan, the developer implements it.

python main.py -p ./projects/app -g "Add OAuth authentication"

Mode 2 — You provide the plan

You describe the architecture and logic in a file, agents implement it directly.

python main.py --plan plan.md -p ./projects/app -g "Implement JWT auth"

See GUIDE.md for detailed instructions on writing plan files.

CLI Options

--project / -p   Path to project directory
--goal    / -g   Goal / task for the agent crew
--tasks   / -t   Comma-separated task list (default: analyze,implement,review)
--plan           Path to plan file with architecture/logic (skips architect)
--memory         Enable agent memory across sessions (ChromaDB)
--human-review   Pause for human approval after every task (overrides tasks.yaml)
--no-human-review  Skip all human approval pauses (overrides tasks.yaml)
--dry-run        Show execution plan without calling API (debug)

Available Tasks

Task	Agent	What it does
analyze_requirements	Architect	Inspects project, creates plan
implement_solution	Senior Dev	Writes code per architect's plan
execute_plan	Senior Dev	Writes code per your plan (--plan)
review_code	Reviewer	Checks quality, security
write_tests	QA Engineer	Writes pytest tests
write_documentation	Tech Writer	Documentation and README
validate_environment	QA Engineer	Checks imports vs requirements.txt

Examples

# Agents build plan and implement (default)
python main.py -p ./projects/my-app -g "Add OAuth via Google"

# Interactive wizard — no flags needed
python main.py

# Your plan — agents implement directly
python main.py --plan plan.md -p ./projects/my-app -g "Implement JWT auth"

# Your plan + full cycle
python main.py --plan plan.md -p ./projects/app -g "Goal" \
  -t execute_plan,review_code,write_tests,validate_environment

# Check what will run (no API calls)
python main.py --dry-run -p ./projects/my-app -g "Test"

# Full development cycle
python main.py -p ./projects/my-app -g "Add OAuth via Google" \
  --tasks analyze_requirements,implement_solution,review_code,write_tests,write_documentation

# With memory — agents remember previous sessions
python main.py -p ./projects/my-app -g "Refactor users module" --memory

# Risky change — pause for human approval after every step
python main.py -p ./projects/my-app -g "Migrate database schema" --human-review

# Fully autonomous pipeline (ignore human_input in tasks.yaml)
python main.py -p ./projects/my-app -g "Fix typo in README" --no-human-review

# Review only
python main.py -p ./projects/legacy -g "Check API security" --tasks review_code

Adding a New Agent

Edit config/agents.yaml:

security_expert:
  role: Security Expert
  goal: Find OWASP Top 10 vulnerabilities and suggest fixes
  backstory: Experienced pentester. Thinks like an attacker, defends like an engineer.
  model_tier: smart
  verbose: true
  allow_delegation: false
  max_iter: 5

Then add a task in config/tasks.yaml — no code changes needed.

Models — Switch Provider in .env, Not YAML

Agents use model tiers (smart / lite) instead of hardcoded model names. Switch provider by editing .env — agents.yaml stays untouched:

# .env — just change these two lines to switch provider
MODEL_SMART=anthropic/claude-sonnet-4-6
MODEL_LITE=anthropic/claude-haiku-4-5

Tier	Used by	Purpose
smart	architect, senior_developer	Complex tasks: planning, code generation
lite	code_reviewer, qa_engineer, technical_writer, devops_engineer	Mechanical tasks: review, tests, docs

Provider examples (copy to .env)

# Anthropic
MODEL_SMART=anthropic/claude-sonnet-4-6
MODEL_LITE=anthropic/claude-haiku-4-5

# OpenAI
MODEL_SMART=openai/gpt-4o
MODEL_LITE=openai/gpt-4o-mini

# Groq (free)
MODEL_SMART=groq/llama-3.3-70b-versatile
MODEL_LITE=groq/llama-3.1-8b-instant

# OpenRouter (free)
MODEL_SMART=openrouter/nvidia/nemotron-3-super-120b-a12b:free
MODEL_LITE=openrouter/nvidia/nemotron-nano-9b-v2:free

# Ollama (local, free)
MODEL_SMART=ollama/qwen3-coder:30b   # or any model from `ollama list`
MODEL_LITE=ollama/llama3.2:3b
OLLAMA_BASE_URL=http://127.0.0.1:11434  # 127.0.0.1 required (macOS IPv6 quirk)

Security

All file tools are sandboxed — agents can only read/write within the --project directory. Path traversal attempts are blocked automatically.

Requirements

• Python 3.11+
• CrewAI 1.12+
• API key in .env (Anthropic, OpenAI, Groq, OpenRouter, or local Ollama)

---

Русская версия

CrewLine — AI-команда разработки: настрой агентов в YAML, укажи на проект, получи код + ревью + тесты.

Быстрый старт

# 1. Установите зависимости
pip install -r requirements.txt

# 2. Создайте .env с API ключом
cp .env.example .env
# Добавьте ANTHROPIC_API_KEY в .env

# 3. Запустите агентов на проекте
python main.py --project ./projects/example_project --goal "Добавить JWT авторизацию"

Два режима работы

Режим 1 — Агенты сами строят план (по умолчанию)

python main.py -p ./projects/app -g "Добавить авторизацию"

Цепочка: архитектор (план) → разработчик (код) → ревьюер (проверка).

Режим 2 — Вы даёте готовый план

python main.py --plan plan.md -p ./projects/app -g "Реализовать JWT авторизацию"

Подробные инструкции — в GUIDE.md.

Параметры CLI

--project / -p   Путь к папке проекта
--goal    / -g   Цель для команды агентов
--tasks   / -t   Список задач через запятую (default: analyze,implement,review)
--plan           Путь к файлу с планом (пропускает архитектора)
--memory         Память агентов между сессиями (ChromaDB)
--human-review   Пауза для ручного подтверждения после каждой задачи (переопределяет tasks.yaml)
--no-human-review  Отключить все паузы (переопределяет tasks.yaml)
--dry-run        Показать план без вызова API (отладка)

Примеры

# Агенты строят план и реализуют
python main.py -p ./projects/my-app -g "Добавить OAuth через Google"

# Готовый план — агенты реализуют напрямую
python main.py --plan plan.md -p ./projects/my-app -g "Реализовать JWT авторизацию"

# Проверить что запустится (без API)
python main.py --dry-run -p ./projects/my-app -g "Тест"

# Рисковое изменение — пауза для ревью после каждого шага
python main.py -p ./projects/my-app -g "Мигрировать БД" --human-review

# Полный цикл
python main.py -p ./projects/my-app -g "Добавить OAuth" \
  --tasks analyze_requirements,implement_solution,review_code,write_tests,write_documentation

Безопасность

Все файловые инструменты работают в sandbox — агенты могут читать/писать только внутри директории --project.

Требования

• Python 3.11+
• CrewAI 1.12+
• API ключ в .env (Anthropic, OpenAI, Groq, OpenRouter или локальный Ollama)

Смена провайдера

Не нужно редактировать agents.yaml. Поменяйте 2 строки в .env:

MODEL_SMART=openai/gpt-4o        # архитектор + разработчик
MODEL_LITE=openai/gpt-4o-mini    # ревьюер + QA + документация