> ## Documentation Index
> Fetch the complete documentation index at: https://apidoc.cometapi.com/llms.txt
> Use this file to discover all available pages before exploring further.

# 将 CrewAI 与 CometAPI 配合使用

> 在一个顺序执行的多智能体 Crew 中，将四种 CrewAI LLM 传输配置与 CometAPI 配合使用。

[CrewAI](https://docs.crewai.com/) 是一个用于协调 AI 智能体和任务的 Python 框架。本指南将配置一个顺序执行的 Crew，通过四种 CrewAI 传输配置调用 CometAPI：OpenAI 聊天补全、OpenAI 响应、Anthropic 消息和 Gemini `generateContent`。

## 前置条件

* Python 3.10–3.13
* [`uv`](https://docs.astral.sh/uv/getting-started/installation/)
* 拥有有效 API 密钥的 CometAPI 账户—— [在控制台中获取您的密钥](https://www.cometapi.com/console/token)
* 适用于下方所示传输方式的四个文本模型 ID

## 配置集成

<Steps>
  <Step title="创建项目并安装 CrewAI">
    创建一个新的 `uv` 项目，并安装包含 Anthropic 和 Google Gen AI 扩展的 CrewAI：

    ```bash theme={null}
    uv init crewai-cometapi
    cd crewai-cometapi
    uv add 'crewai[anthropic,google-genai]'
    ```
  </Step>

  <Step title="设置 API 密钥和模型 ID">
    设置 CometAPI API 密钥，并为每种传输方式设置一个模型 ID：

    ```bash theme={null}
    read -rsp "CometAPI API key: " COMETAPI_KEY
    printf '\n'
    export COMETAPI_KEY

    export COMETAPI_CHAT_MODEL_ID=your-model-id
    export COMETAPI_RESPONSES_MODEL_ID=your-model-id
    export COMETAPI_ANTHROPIC_MODEL_ID=your-model-id
    export COMETAPI_GEMINI_MODEL_ID=your-model-id
    ```

    将每个 `your-model-id` 值替换为 [CometAPI 模型页面](/zh-Hans/overview/models)中的模型 ID。四个值可以不同。将每个模型 ID 分配给其对应的传输方式。
  </Step>

  <Step title="创建顺序执行的 Crew">
    将以下示例保存为 `crew.py`：

    ```python theme={null}
    import os

    from crewai import Agent, Crew, LLM, Process, Task


    api_key = os.environ["COMETAPI_KEY"]

    chat_llm = LLM(
        model=os.environ["COMETAPI_CHAT_MODEL_ID"],
        custom_openai=True,
        api="completions",
        base_url="https://api.cometapi.com/v1",
        api_key=api_key,
        max_tokens=512,
        max_retries=0,
    )

    responses_llm = LLM(
        model=os.environ["COMETAPI_RESPONSES_MODEL_ID"],
        custom_openai=True,
        api="responses",
        base_url="https://api.cometapi.com/v1",
        api_key=api_key,
        max_completion_tokens=512,
        max_retries=0,
    )

    anthropic_llm = LLM(
        model=f"anthropic/{os.environ['COMETAPI_ANTHROPIC_MODEL_ID']}",
        base_url="https://api.cometapi.com",
        api_key=api_key,
        max_tokens=512,
        max_retries=0,
    )

    gemini_llm = LLM(
        model=f"gemini/{os.environ['COMETAPI_GEMINI_MODEL_ID']}",
        api_key=api_key,
        max_output_tokens=512,
        thinking_config={"thinking_budget": 0, "include_thoughts": False},
        client_params={
            "http_options": {
                "base_url": "https://api.cometapi.com",
                "api_version": "v1beta",
            }
        },
    )

    audience_agent = Agent(
        role="Audience researcher",
        goal="Define the audience for an API migration brief",
        backstory="You turn product goals into a precise audience statement.",
        llm=chat_llm,
        tools=[],
        allow_delegation=False,
        max_iter=1,
        max_retry_limit=0,
        verbose=False,
    )

    requirements_agent = Agent(
        role="Requirements planner",
        goal="Turn an audience statement into implementation requirements",
        backstory="You write concise, testable requirements for API teams.",
        llm=responses_llm,
        tools=[],
        allow_delegation=False,
        max_iter=1,
        max_retry_limit=0,
        verbose=False,
    )

    risk_agent = Agent(
        role="Risk reviewer",
        goal="Identify the most important migration risk",
        backstory="You review plans for practical delivery risks.",
        llm=anthropic_llm,
        tools=[],
        allow_delegation=False,
        max_iter=1,
        max_retry_limit=0,
        verbose=False,
    )

    editor_agent = Agent(
        role="Release brief editor",
        goal="Combine research, requirements, and risk into one brief",
        backstory="You preserve source findings while producing clear summaries.",
        llm=gemini_llm,
        tools=[],
        allow_delegation=False,
        max_iter=1,
        max_retry_limit=0,
        verbose=False,
    )

    audience_task = Task(
        description=(
            "Define one target audience for a team moving an existing OpenAI "
            "integration to a multi-provider API. Return only an Audience heading "
            "and one sentence of at most 20 words."
        ),
        expected_output="An Audience heading followed by one concise sentence.",
        agent=audience_agent,
    )

    requirements_task = Task(
        description=(
            "Using the audience statement in your context, return only a "
            "Requirements heading and exactly two numbered requirements. Keep each "
            "requirement to at most 18 words."
        ),
        expected_output="A Requirements heading followed by two numbered items.",
        agent=requirements_agent,
        context=[audience_task],
    )

    risk_task = Task(
        description=(
            "Using the audience and requirements in your context, return only a "
            "Risk heading, one risk sentence, and one mitigation sentence. Keep "
            "each sentence to at most 18 words."
        ),
        expected_output="A Risk heading with one risk and one mitigation.",
        agent=risk_agent,
        context=[audience_task, requirements_task],
    )

    brief_task = Task(
        description=(
            "Create a release brief from all prior task outputs. Return only these "
            "sections: Audience with one sentence; Requirements with two numbered "
            "items; Risk with one risk and one mitigation. Keep the exact headings "
            "Audience, Requirements, and Risk."
        ),
        expected_output=(
            "A concise release brief with Audience, Requirements, and Risk headings."
        ),
        agent=editor_agent,
        context=[audience_task, requirements_task, risk_task],
    )

    crew = Crew(
        agents=[
            audience_agent,
            requirements_agent,
            risk_agent,
            editor_agent,
        ],
        tasks=[audience_task, requirements_task, risk_task, brief_task],
        process=Process.sequential,
        verbose=False,
    )

    result = crew.kickoff()

    for index, task_output in enumerate(result.tasks_output, start=1):
        print(f"\n--- Task {index} ---\n{task_output.raw}")
    ```

    这些智能体不使用工具、不能委派任务，并且一次运行一个。第一个任务之后的每个任务都会在 `context` 中声明其前置任务，因此 CrewAI 会将这些先前的输出包含在下一个任务中。
  </Step>

  <Step title="运行 Crew">
    在同一 shell 会话中运行示例：

    ```bash theme={null}
    uv run python crew.py
    ```

    脚本会输出全部四个任务的结果，包括最终的发布摘要。
  </Step>
</Steps>

## 路由映射

| CrewAI 配置                                    | CometAPI 路由                                   | 基础 URL                        |
| -------------------------------------------- | --------------------------------------------- | ----------------------------- |
| `custom_openai=True`, `api="completions"`    | `POST /v1/chat/completions`                   | `https://api.cometapi.com/v1` |
| `custom_openai=True`, `api="responses"`      | `POST /v1/responses`                          | `https://api.cometapi.com/v1` |
| `anthropic/` 模型前缀                            | `POST /v1/messages`                           | `https://api.cometapi.com`    |
| `gemini/` 模型前缀和 `client_params.http_options` | `POST /v1beta/models/{model}:generateContent` | `https://api.cometapi.com`    |

上述两种 OpenAI 配置需要在 `/v1` 中使用 `base_url` 后缀。Anthropic 和 Gemini 配置会添加各自带版本的路由，因此其基础 URL 是不含 `/v1` 的 CometAPI 源站地址。

## 选择模型 ID

使用 [CometAPI 模型页面](/zh-Hans/overview/models) 为每种传输方式选择一个文本模型 ID。将模型 ID 存储在环境变量中，以便无需编辑 Python 文件即可更改它们。请勿为环境变量添加提供商前缀。示例会在 CrewAI 使用这些前缀进行提供商选择的位置添加 `anthropic/` 和 `gemini/`。

## 故障排除

<AccordionGroup>
  <Accordion title="Anthropic 或 Gemini 提供商导入失败">
    在项目中运行 `uv add 'crewai[anthropic,google-genai]'`。基础 CrewAI 包默认不会安装这两个可选的提供商 SDK。
  </Accordion>

  <Accordion title="响应请求返回参数错误">
    移除不属于所选模型接口的可选参数。该示例仅为响应配置了输出 Token 上限。
  </Accordion>

  <Accordion title="请求使用了错误的路由">
    将 `LLM` 配置与路由映射表进行比较。尤其要使用具有正确 `custom_openai=True` 值的 `api`，并将 Gemini 基础 URL 保持在 `client_params.http_options` 内。
  </Accordion>
</AccordionGroup>

## 相关资源

* [CrewAI 1.15.12 LLM 文档](https://docs.crewai.com/v1.15.12/en/concepts/llms)
* [CrewAI 智能体](https://docs.crewai.com/v1.15.12/en/concepts/agents)
* [CrewAI 任务](https://docs.crewai.com/v1.15.12/en/concepts/tasks)
* [CrewAI Crew](https://docs.crewai.com/v1.15.12/en/concepts/crews)
* [CometAPI 聊天补全](/zh-Hans/api/text/chat)
* [CometAPI 响应](/zh-Hans/api/text/responses)
* [CometAPI Anthropic 消息](/zh-Hans/api/text/anthropic-messages)
* [CometAPI Gemini generateContent](/zh-Hans/api/text/gemini-generating-content)

<script type="application/ld+json">
  {`
    {
    "@context": "https://schema.org",
    "@graph": [
      {
        "@type": "HowTo",
        "@id": "https://apidoc.cometapi.com/integrations/crewai#howto",
        "name": "将 CrewAI 与 CometAPI 配合使用",
        "description": "在一个顺序执行的多智能体 Crew 中，将四种 CrewAI LLM 传输配置与 CometAPI 配合使用。",
        "step": [
          {
            "@type": "HowToStep",
            "@id": "https://apidoc.cometapi.com/integrations/crewai#step-1",
            "position": 1,
            "name": "创建项目并安装 CrewAI",
            "text": "创建一个 uv 项目，并安装包含 Anthropic 和 Google Gen AI 扩展的 CrewAI。"
          },
          {
            "@type": "HowToStep",
            "@id": "https://apidoc.cometapi.com/integrations/crewai#step-2",
            "position": 2,
            "name": "设置 API 密钥和模型 ID",
            "text": "设置 CometAPI API 密钥，并为每种传输方式设置一个当前模型 ID。"
          },
          {
            "@type": "HowToStep",
            "@id": "https://apidoc.cometapi.com/integrations/crewai#step-3",
            "position": 3,
            "name": "创建顺序执行的 Crew",
            "text": "在一个顺序执行的 Crew 中配置四个 LLM 适配器、四个智能体和四个通过上下文关联的任务。"
          },
          {
            "@type": "HowToStep",
            "@id": "https://apidoc.cometapi.com/integrations/crewai#step-4",
            "position": 4,
            "name": "运行 Crew",
            "text": "运行 Crew 并检查全部四个任务输出。"
          }
        ]
      },
      {
        "@type": "BreadcrumbList",
        "itemListElement": [
          {
            "@type": "ListItem",
            "position": 1,
            "name": "CometAPI 文档",
            "item": "https://apidoc.cometapi.com/"
          },
          {
            "@type": "ListItem",
            "position": 2,
            "name": "集成",
            "item": "https://apidoc.cometapi.com/integrations"
          },
          {
            "@type": "ListItem",
            "position": 3,
            "name": "将 CrewAI 与 CometAPI 配合使用",
            "item": "https://apidoc.cometapi.com/integrations/crewai"
          }
        ]
      }
    ]
    }
    `}
</script>