菜单
Use when the user wants to give a cuga agent a new capability via a Python function, REST/OpenAPI service, or MCP server, e.g. "add a tool", "connect this API to cuga", "register an MCP server".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when the user wants to govern agent behavior with a cuga policy - block/redirect an intent, add step-by-step playbook guidance, require approval before a tool runs, enhance a tool's description, or reshape agent output format.
Use when the user wants to write Python code that creates or invokes a CugaAgent or CugaSupervisor, e.g. "build an agent with cuga", "how do I call agent.invoke", "set up a multi-agent supervisor".
Use when the user wants a running CugaAgent to load a new capability at runtime via cuga's own skill system, e.g. "add a skill to my agent", "create a .cuga/skills SKILL.md", "how does load_skill work".
Use when the user wants to install, launch, or build something with the cuga agent framework (pip package `cuga`), or asks "how do I use cuga" / "how do I build an agent with cuga" — routes to the right cuga-harness-kit skill instead of guessing.
Use when the user wants to install cuga, start the demo/web UI, or asks what `cuga start` does or which service/mode to run.
Use when the user wants a cuga agent to ingest, search, or answer questions from documents (PDF/DOCX/XLSX/PPTX/HTML/Markdown/images) - RAG / knowledge base features.
| name | cuga-build-tool |
| description | Use when the user wants to give a cuga agent a new capability via a Python function, REST/OpenAPI service, or MCP server, e.g. "add a tool", "connect this API to cuga", "register an MCP server". |
cuga supports three tool integration types:
| Type | Best for | Configured via | Loading |
|---|---|---|---|
| LangChain | Python functions, rapid prototyping | direct import, pass to CugaAgent(tools=[...]) | Runtime |
| OpenAPI | REST APIs, existing services | mcp_servers.yaml | Build-time |
| MCP | Custom protocols, complex integrations | mcp_servers.yaml | Build-time |
from langchain_core.tools import tool
@tool
def lookup_order(order_id: str) -> str:
"""Look up an order by ID and return its status."""
...
agent = CugaAgent(tools=[lookup_order])
This is a runtime-loaded tool: no config file, no restart needed, just pass it into the CugaAgent constructor. Write a clear docstring — it becomes the tool description the agent's reasoning engine sees.
For REST APIs or MCP servers, add an entry to mcp_servers.yaml (path: src/cuga/backend/tools_env/registry/config/mcp_servers.yaml in a cuga-agent checkout). These are build-time: the tool registry (cuga start registry, or bundled into any demo_*/manager service) loads this config at startup, so changes need a restart to take effect.
See the registry's own README for the exact config schema and worked examples: src/cuga/backend/tools_env/registry/README.md, plus docs/examples/cuga_with_runtime_tools/README.md for a full walkthrough combining different tool types with MCP.
mcp_servers.yaml.mcp_servers.yaml.Once a tool exists, you can shape how the agent uses it without touching its code — that's a tool_guide or tool_approval policy (require confirmation before a sensitive tool runs, or append extra usage guidance/examples to its description). See cuga-author-policy.