camel-ai/owl

Purpose and Entry Points

OWL is a Python framework for real-world task automation through collaborating LLM agents, built on the external CAMEL-AI framework. Primary CLI examples construct and run an agent “society”; the local Gradio UI exposes the same workflow in English, Chinese, and Japanese.

• examples/run.py

• owl/webapp.py

• owl/webapp_zh.py

Workforce Orchestration

The core execution model is enhanced role-playing: agents exchange messages, select tools, process results, and converge on an answer plus chat history/token usage. The implementation is thin in this repository and delegates foundational agent, model, and toolkit behavior to the installed CAMEL dependency.

• owl/utils/enhanced_role_playing.py

• examples/run.py

• pyproject.toml

Task Input and Document Boundary

User tasks may include local file paths. Document utilities provide the repository-local adaptation layer for extracting or preparing document content, while tool selection and model-driven invocation remain CAMEL responsibilities.

• owl/utils/document_toolkit.py

• owl/utils/common.py

• examples/run.py

Tool-Enabled Execution

Agents operate through configured toolkits: search, browser automation, multimodal analysis, document parsing, code execution, filesystem writing, and MCP services. Tool calls cross the main trust/runtime boundary: local subprocesses and browsers, remote search/content providers, and optional MCP servers.

• examples/run.py

• community_usecase/qwen3_mcp/mcp_sse_config.json

• community_usecase/Puppeteer MCP/mcp_servers_config.json

Model and External-Service Integration

The framework requires tool-calling LLMs and uses environment-provided credentials. It supports OpenAI and other CAMEL-supported model backends; task capabilities can additionally depend on OpenAI audio services, search engines, Playwright/browser channels, Node.js MCP services, and remote SSE MCP endpoints.

• owl/.env_template

• examples/run_claude.py

• examples/run_gemini.py

Web Interaction Surface

The web applications are local Gradio front ends for model selection, environment configuration, chat submission, and task history. They are alternate entry points rather than a separate server-side product; configured model APIs are the principal external data egress.

• owl/webapp.py

• owl/webapp_jp.py

• owl/webapp_backup.py

Deployment and Runtime

Supported runtime is Python 3.10–3.12, installed via package metadata or requirements. Docker Compose packages the environment, including browser-oriented execution requirements; Playwright may need OS dependencies and virtual display support.

• pyproject.toml

• requirements.txt

• .container/docker-compose.yml

• .container/Dockerfile

Extensions and Reference Implementations

community_usecase/ contains independent vertical applications and MCP demonstrations, not core OWL runtime modules. They illustrate how the orchestration/tool boundary is reused for domain workflows such as investment analysis, interview preparation, and document/resume processing.

• community_usecase/a_share_investment_agent_camel/src/main.py

• community_usecase/stock-analysis/run.py

• community_usecase/OWL Interview Preparation Assistant/main.py