tools-runtimes-and-services

Tools, Runtimes, and Services

Use this sub-skill when a task needs CAMEL agents to call tools, expose tools through services, or execute code in a controlled runtime. It covers camel.toolkits, FunctionTool, toolkit get_tools() patterns, MCPToolkit, OpenAPIToolkit, CodeExecutionToolkit, TerminalToolkit, camel.runtimes, camel.interpreters, CAMEL agent MCP service patterns, and camel.services.agent_openapi_server.

Do not use this sub-skill for provider/model credentials or model backend selection; route that to the model/provider sub-skill. Do not use it for retrieval-specific memory/RAG tools beyond high-level routing; route retrieval work to ../memory-rag-and-data/SKILL.md when available. Do not use it for multi-agent loop design except to show how the agent receives tools.

Fast Routing

• For custom Python functions and tool schemas, read references/api-reference.md and optionally run scripts/inspect_tool_schema.py.
• For selecting a built-in toolkit or optional extra, read references/toolkit-catalog.md.
• For MCP, OpenAPI, terminal/code execution, runtimes, interpreters, AgentOpenAPI, or service deployment patterns, read references/runtime-service-patterns.md.
• For failures around schemas, external_tools, missing extras, API credentials, MCP startup, Docker/browser/runtime availability, sandbox safety, or OpenAPI auth, read references/troubleshooting.md.

Minimal Patterns

from camel.agents import ChatAgent
from camel.toolkits import FunctionTool, SearchToolkit

def add(a: int, b: int) -> int:
    """Add two integers."""
    return a + b

agent = ChatAgent(tools=[FunctionTool(add), *SearchToolkit().get_tools()])

from camel.toolkits import MCPToolkit

async with MCPToolkit(config_dict={"mcpServers": {"demo": {"command": "python", "args": ["server.py"]}}}) as toolkit:
    agent = ChatAgent(tools=toolkit.get_tools())

Safety Defaults

Prefer pure schema inspection before live calls, treat browser/Docker/terminal/runtime tools as side-effecting, and keep secrets in environment variables or service config outside skill content. For code execution, start with internal_python or a sandboxed runtime, explicitly decide require_confirm, and never enable unsafe_mode=True for untrusted model-generated code without an external sandbox and human approval.