今天的羊毛复盘
我花了 1.2 个番茄钟,薅到了 OpenRouter 的 5 个免费模型
不是教程党式搬运,是我今天自己跑通的一条可直接接进 Agent 的路径。
今天又去薅了几个免费的羊毛。
准确一点说,我花了 1.2 个番茄钟,把 OpenRouter 的 5 个免费模型接进了自己的 Agent 里。
这件事为什么值得记一下?
因为它不是那种“薅完就结束”的羊毛。
一旦你把它接进自己的 Agent,以后写内容、跑任务、做一点 Vibe Coding,手里就多了一条备用算力链路。
今天我先把路径跑通,具体使用感受后面再发。这篇先把最重要的部分记下来:怎么拿到、怎么接入、哪里容易踩坑。
先说结论:这事不复杂,真正卡人的不是操作本身,而是你有没有把要给 Agent 的信息一次性说清楚。
01 第一步:去 OpenRouter 注册,然后创建 API Key
第一步很直接。
先去 openrouter.ai 注册账号,然后进入 API 密钥管理页面,点一下“创建新密钥”。
密钥生成后大概长这样:
sk-or-XXXX
这一步有两个坑必须注意。
坑 1:密钥只在创建时可见一次。复制完就找不回来了,丢了只能重新生成。
坑 2:如果你后面是让 Agent 帮你接入,不要只说“帮我配 OpenRouter”,要把密钥、目标 Agent、模型名单、以及你想做内容创作还是 Vibe Coding 一次性交代清楚。
02 第二步:接入你的 Agent,CC 和 Hermes 都可以
第二步就是接入你自己的 Agent。
CC 还是 Hermes 都行,这个看你平时用哪个更顺手。
如果接入 CC,可用 CC switch 一键切换。
我这次选择接入 Hermes,先做了最基础的一步:
第 1 步:设置 API Key
hermes config set OPENROUTER_API_KEY <你的OpenRouter密钥>
密钥获取地址就是:https://openrouter.ai/keys
然后可以直接这样启动:
hermes chat --provider openrouter --model 'openrouter/owl-alpha:free'
我后面其实没有自己一条条手敲完,而是让 CC 帮我继续配,把需要的信息一次性给它,让它帮我找出 5 个免费模型,再把配置补全。
这 5 个免费模型,我最后留下的是这套组合。
03 我最后留下的 5 个免费模型
这次我整理出来的 5 个免费模型,大概可以这么理解:
openrouter/owl-alpha:通用 Agent 任务,综合最强。
nvidia/nemotron-3-ultra-550b-a55b:free:长推理、复杂任务,1M 超长上下文。
google/gemma-4-31b-it:free:多模态 + 创意内容。
openai/gpt-oss-120b:free:通用长文本,能力均衡。
poolside/laguna-m.1:free:编码专用,Vibe Coding 首选。
如果你只是想有一条备用算力链路,这 5 个已经够你搭出一个还不错的最小组合了。
04 内容创作配置,我是这么配的
我这次做了两套配置。
第一套偏内容创作,逻辑很简单:默认走 owl-alpha,如果断了,就自动切到后面的 fallback 模型。
model: default: openrouter/owl-alpha provider: openrouter base_url: https://openrouter.ai/api/v1 fallback_providers: - provider: openrouter model: nvidia/nemotron-3-ultra-550b-a55b - provider: openrouter model: google/gemma-4-31b-it - provider: openrouter model: openai/gpt-oss-120b - provider: openrouter model: poolside/laguna-m.1 provider_routing: sort: throughput data_collection: deny
这个配置的核心点就两个:
第一:默认模型先扛住日常任务。
第二:如果默认模型断掉,后面的模型自动接上,不用你自己再切来切去。
05 如果是 Vibe Coding,我会优先这么配
第二套其实思路差不多,只是我的重点不一样。
如果是偏 Vibe Coding 的场景,我会更关注编码模型的手感和切换稳定性,所以我会把 poolside/laguna-m.1 放进可用链路里,作为编码向的优先备胎。
我这次跑通后,Hermes 这边已经能正常切换,配置算是成功了。
配置成功之后,我这边切到 Laguna 是能正常工作的。
06 这件事真正省时间的地方
我这次的感受很明确:
真正省时间的,不是你手敲那几行命令。
真正省时间的是,你有没有一次性把下面这些信息给到 Agent:
1. 你要接的是 CC 还是 Hermes。
2. 你的 OpenRouter Key 是什么。
3. 你要找的是“免费模型”,不是泛泛推荐。
4. 你是拿它做内容创作,还是拿它做 Vibe Coding。
5. 你希望它断了自动切,而不是每次手动换模型。
你把这些信息一口气说清楚,Agent 干活就很快。
你如果只说一句“帮我接一下 OpenRouter”,它大概率会来回追问,或者配出一套不贴你实际场景的东西。
我今天给自己的记录
这 1.2 个番茄钟没有白花。不是因为省了多少钱,而是因为我又多打通了一条备用算力链路。后面写内容、跑任务、做编码实验时,手里就多了一张牌。
今天先把路径记到这里。后续我再单独发一下这 5 个免费模型的实用感受,到底行不行
Today's Woolly Recap
I spent 1.2 pomodoros hooking up 5 free models from OpenRouter to my Agent.
This isn't a tutorial dump — it's my actual, working setup from today, ready to plug into your Agent.
Today I grabbed a few more freebies.
Strictly speaking, I spent 1.2 pomodoros integrating 5 free OpenRouter models into my own Agent setup.
Why is this worth noting?
Because it's not a "use once and done" kind of deal.
Once you've got it hooked into your Agent, you've always got a backup compute path for content writing, running tasks, or even a bit of vibe coding.
Today I'm just mapping out the path. I'll share the actual performance later. This post is about the essentials: how to get it, how to hook it up, and where the traps lie.
Bottom line: This isn't rocket science. The real bottleneck isn't the tech — it's whether you're communicating all the necessary info to your Agent in one shot.
01 First Step: Register on OpenRouter and Create an API Key
Step one is straightforward.
Head over to openrouter.ai, sign up, and go to the API key management page to create a new key.
The key will look something like this:
sk-or-XXXX
There are two critical pitfalls to avoid here.
Pitfall 1: The key is only visible once — during creation. If you lose it, you're out of luck until you generate a new one.
Pitfall 2: When instructing your Agent, don't just say "configure OpenRouter." You need to specify the API key, the target Agent, the model list, and the intended use case (e.g., content creation or vibe coding).
02 Second Step: Hooking Up to Your Agent (CC or Hermes)
Now it's time to connect to your Agent — be it CC or Hermes.
If you're using CC, you can switch easily with the CC switch command.
I'm using Hermes here, so the first step is simple:
Step 1: Set the API Key
You can find the key at: https://openrouter.ai/keys
Then start the integration like this:
Honestly, I didn't type all that out. I told CC what I needed, and it figured out the 5 best free models and completed the config for me.
03 My Final List of 5 Free Models
Here's the breakdown of my final selection:
openrouter/owl-alpha: The Swiss Army knife — best for general Agent tasks.
nvidia/nemotron-3-ultra-550b-a55b:free: The thinker — for complex reasoning and tasks that require a 1M+ context window.
google/gemma-4-31b-it:free: The creator — excels at multimodal content and creative writing.
openai/gpt-oss-120b:free: The balanced reader — solid for general long-text comprehension.
poolside/laguna-m.1:free: The coder — specifically tuned for vibe coding scenarios.
This mix should cover the basics for a minimal backup setup.
04 My Content Creation Configuration
I set up two distinct workflows.
The first is for content creation. The logic is simple: start with owl-alpha, and if it fails, automatically switch to the fallback models.
Two key points here:
First, the default model has to be robust enough for daily use.
Second, the fallback mechanism needs to be automatic — no manual switching required if the primary mode goes down.
05 My Vibe Coding Configuration
The second setup is for Vibe Coding.
For this, I focus on the models' "feel" and reliability. I prioritize poolside/laguna-m.1 as a dedicated coding backup, as it's specifically tuned for that domain.
The good news is, after this setup, Hermes is running smoothly, and transitions are working as expected.
06 Where This Really Saves Time
The takeaway here isn't about saving money on tokens.
The real time-saver is in how you communicate with your Agent. If you tell it exactly what you need — the agent type, the API key, the model criteria, and the use case — it can automate the entire selection and configuration process.
Vagueness leads to wasted time. Clarity lets the AI do the heavy lifting.
07 Final Thoughts
Those 1.2 pomodoros were well-spent. I've added a crucial layer of redundancy to my workflow. Now, if my primary cloud provider has a hiccup, I've got a viable alternative ready to go for content or coding tasks.
2. What is your OpenRouter Key?
3. You are looking for "free models," not a generic recommendation.
4. What will you use it for? Content creation or Vibe Coding?
5. You want seamless switching without manually changing models each time.
Clearly state these details, and the agent will quickly provide a solution.
If you only say "Connect me to OpenRouter," it will likely ask questions back and forth or recommend something unsuitable.
My notes for today:
These 1.2 pomodoro timers are well-spent. Not just for saving money, but for establishing a backup computing resource path. In the future, whether for writing content, running tasks, or coding experiments, I have one more option at hand.
Record the path here today. Next, I'll share my experience with the 5 free models separately — do they work?

