Launch HN: Dedalus Labs (YC S25) – 代理人的 Vercel
4 分•作者: windsor•9 个月前
嘿，HN！我们是 Dedalus Labs 的 Windsor 和 Cathy（<a href="https://www.dedaluslabs.ai/" rel="nofollow">https://www.dedaluslabs.ai/</a>），一个供开发者构建智能体 AI 应用的云平台。我们的 SDK 允许你将任何 LLM 连接到任何 MCP 工具——本地的或由我们托管的。无需 Dockerfile 或 YAML 配置。
这里有一个演示：<a href="https://youtu.be/s2khf1Monho?si=yiWnZh5OP4HQcAwL&t=11" rel="nofollow">https://youtu.be/s2khf1Monho?si=yiWnZh5OP4HQcAwL&t=11</a>
去年十月，我（Windsor）试图在云端构建一个 LLM 可以进行工具调用的有状态代码执行沙盒。那是在 MCP 发布之前，可以说构建过程非常令人恼火……我一直在想，“为什么我不能直接将 `tools=code_execution` 传递给模型，然后让它……工作呢？”
即使有了 MCP，你仍然需要运行本地服务器，并在 OpenAI、Anthropic、Google 等之间手动连接 API 认证和格式化，然后才能发布任何东西。每次更改都意味着重新部署、网络配置，以及浪费数小时来处理 AWS。在构建产品时，你可不想花几个小时阅读文档和与云设置斗争！
Dedalus 将此简化为仅一个 API 端点，因此过去需要 2 周的设置时间现在只需 5 分钟。我们允许你将可流式 HTTP MCP 服务器上传到我们的平台。部署后，我们提供与 OpenAI 兼容的 SDK，你可以将其放入你的代码库中使用 MCP 驱动的 LLM。我们的目标是让任何人、在任何地方都能为他们的 LLM 配备强大的工具，用于函数调用。
你编写的代码大致如下：
```python
client = Dedalus()
runner = DedalusRunner(client)
result = runner.run(
input=prompt,
tools=[tool_1, tool_2],
mcp_servers=["author/server-1”, “author/server-2”],
model=["openai/gpt-4.1”, “anthropic/claude-sonnet-4-20250514”], # 默认为列表中的第一个模型
stream=True,
)
stream_sync(result) # 流式传输结果，也支持工具调用
```
我们的文档从 <a href="https://docs.dedaluslabs.ai" rel="nofollow">https://docs.dedaluslabs.ai</a> 开始。这里有一个简单的 Hello World 示例：<a href="https://docs.dedaluslabs.ai/examples/01-hello-world" rel="nofollow">https://docs.dedaluslabs.ai/examples/01-hello-world</a>。对于基本的工具执行，请参阅 <a href="https://docs.dedaluslabs.ai/examples/02-basic-tools" rel="nofollow">https://docs.dedaluslabs.ai/examples/02-basic-tools</a>。网站上还有更多示例，包括更复杂的示例，例如使用 Open Meteo MCP 进行天气预报：<a href="https://docs.dedaluslabs.ai/examples/use-case/weather-forecaster" rel="nofollow">https://docs.dedaluslabs.ai/examples/use-case/weather-foreca...</a>。
毫无疑问，MCP 领域仍然存在很多问题。一个大问题是身份验证（我和我的团队开玩笑说 MCP 中的“S”代表“安全”）。目前的 MCP 服务器需要同时充当身份验证服务器和资源服务器。这很难正确实现，对服务器编写者来说要求太高了，也就是说，人们只想公开一个资源端点就完事了。
尽管如此，我们仍然看好 MCP。目前的缺点并非不可挽回，我们预计未来的修订将解决人们目前的大部分疑虑。我们认为有用的 AI 智能体必然会习惯性地调用工具，而 MCP 是为模型配备工具的一个相当不错的方法。
我们还没有完全实现我去年十月想要的有状态代码执行沙盒，但我们正在努力！发布安全且有状态的 MCP 服务器是我们的首要任务，我们将在下个月推出我们的身份验证解决方案。我们还在开发一个 MCP 市场，让人们可以将其工具货币化，同时我们处理计费和收入分成。
我们非常重视开源，目前有以下 SDK（MIT 许可）：
<a href="https://github.com/dedalus-labs/dedalus-sdk-python" rel="nofollow">https://github.com/dedalus-labs/dedalus-sdk-python</a>
<a href="https://github.com/dedalus-labs/dedalus-sdk-typescript" rel="nofollow">https://github.com/dedalus-labs/dedalus-sdk-typescript</a>
<a href="https://github.com/dedalus-labs/dedalus-sdk-go" rel="nofollow">https://github.com/dedalus-labs/dedalus-sdk-go</a>
<a href="https://github.com/dedalus-labs/dedalus-openapi" rel="nofollow">https://github.com/dedalus-labs/dedalus-openapi</a>
我们很乐意收到关于你们认为是什么阻碍你们将 MCP 服务器集成或使用工具调用 LLM 纳入当前工作流程的反馈。
谢谢 HN！
查看原文
Hey HN! We are Windsor and Cathy of Dedalus Labs (<a href="https://www.dedaluslabs.ai/" rel="nofollow">https://www.dedaluslabs.ai/</a>), a cloud platform for developers to build agentic AI applications. Our SDK allows you to connect any LLM to any MCP tools – local or hosted by us. No Dockerfiles or YAML configs required.<p>Here’s a demo: <a href="https://youtu.be/s2khf1Monho?si=yiWnZh5OP4HQcAwL&t=11" rel="nofollow">https://youtu.be/s2khf1Monho?si=yiWnZh5OP4HQcAwL&t=11</a><p>Last October, I (Windsor) was trying to build a stateful code execution sandbox in the cloud that LLMs could tool-call into. This was before MCP was released, and let’s just say it was super annoying to build… I was thinking to myself the entire time “Why can’t I just pass in `tools=code_execution` to the model and just have it…work?<p>Even with MCP, you’re stuck running local servers and handwiring API auth and formatting across OpenAI, Anthropic, Google, etc. before you can ship anything. Every change means redeploys, networking configs, and hours lost wrangling AWS. Hours of reading docs and wrestling with cloud setup is not what you want when building your product!<p>Dedalus simplifies this to just one API endpoint, so what used to take 2 weeks of setup can take 5 minutes. We allow you to upload streamable HTTP MCP servers to our platform. Once deployed, we offer OpenAI-compatible SDKs that you can drop into your codebase to use MCP-powered LLMs. The idea is to let anyone, anywhere, equip their LLMs with powerful tools for function calling.<p>The code you write looks something like this:<p><pre><code> python
client = Dedalus()
runner = DedalusRunner(client)
result = runner.run(
input=prompt,
tools=[tool_1, tool_2],
mcp_servers=["author/server-1”, “author/server-2”],
model=["openai/gpt-4.1”, “anthropic/claude-sonnet-4-20250514”], # Defaults to first model in list
stream=True,
)
stream_sync(result) # Streams result, supports tool calling too
</code></pre>
Our docs start at <a href="https://docs.dedaluslabs.ai" rel="nofollow">https://docs.dedaluslabs.ai</a>. Here’s a simple Hello World example: <a href="https://docs.dedaluslabs.ai/examples/01-hello-world" rel="nofollow">https://docs.dedaluslabs.ai/examples/01-hello-world</a>. For basic tool execution, see <a href="https://docs.dedaluslabs.ai/examples/02-basic-tools" rel="nofollow">https://docs.dedaluslabs.ai/examples/02-basic-tools</a>. There are lots more examples on the site, including more complex ones like using the Open Meteo MCP to do weather forecasts: <a href="https://docs.dedaluslabs.ai/examples/use-case/weather-forecaster" rel="nofollow">https://docs.dedaluslabs.ai/examples/use-case/weather-foreca...</a>.<p>There are still a bunch of issues in the MCP landscape, no doubt. One big one is authentication (my team and I joke that the “S” in MCP stands for “security”). MCP servers right now are expected to act as both the authentication server <i>and</i> the resource server. This is tricky to implement correctly, and it’s too much to ask of server writers, i.e. people just want to expose a resource endpoint and be done.<p>Still, we are bullish on MCP. Current shortcomings are not irrecoverable, and we expect future amendments to resolve most qualms that people currently have. We think that useful AI agents are bound to be habitual tool callers, and MCP is a pretty decent way to equip models with tools.<p>We aren’t <i>quite</i> yet at the stateful code execution sandbox that I wanted last October, but we’re getting there! Shipping secure and stateful MCP servers is high on our priority list, and we’ll be launching our auth solution next month. We’re also working on an MCP marketplace, to let people monetize their tools, while we handle billing and rev-share.<p>We’re big on open sourcing things and have these SDKs so far (MIT licensed):<p><a href="https://github.com/dedalus-labs/dedalus-sdk-python" rel="nofollow">https://github.com/dedalus-labs/dedalus-sdk-python</a><p><a href="https://github.com/dedalus-labs/dedalus-sdk-typescript" rel="nofollow">https://github.com/dedalus-labs/dedalus-sdk-typescript</a><p><a href="https://github.com/dedalus-labs/dedalus-sdk-go" rel="nofollow">https://github.com/dedalus-labs/dedalus-sdk-go</a><p><a href="https://github.com/dedalus-labs/dedalus-openapi" rel="nofollow">https://github.com/dedalus-labs/dedalus-openapi</a><p>We would love feedback on what you guys think are the biggest barriers that keep you from integrating MCP servers or using tool calling LLMs into your current workflow.<p>Thanks HN!