nezhar

Announcement https://hy.tencent.com/research/hy3

Free for two weeks on OpenRouter https://openrouter.ai/tencent/hy3:free

Still looking for this, the OpenKB project look promising.

I agree with the vendor lock-in aspect. My strategy was to utilize multiple agents with different APIs.

you've reached your plan's message limit

For me, building with open weights models sounds like the right approach — you are able to switch providers, and you can control where the server is running.

You don't have any guarantees in terms of data, that's true, you rely on the provider. But this is similar to a database or other services where you don't have the knowledge or resources to run them yourself. Hardware cost is an additional factor here.

If on the other hand your idea works out and the model fits the use case, you can always decide to move to a dedicated infrastructure later.

The same thing can happen in development. Data exfiltration or local file removals are often downplayed; I wonder why nobody talks about the lethal trifecta anymore.

I was just wondering if this might be the first incident. Are there any other public stories available?

It is very scary to consider the consequences that such a transfer can have.

This is, honestly, very sad.

I'm also curious about this one, also I want to compare this to vLLM.

I created VibePod, which allows you to sandbox the agents in containers and monitor their activities. It also supports OpenCode.

https://github.com/VibePod/vibepod-cli

There is also an Auto-reload option, I think it is off by default.

You can use this if you started your Pro or Max subscription before Wednesday, February 4, 2026 at 11:59 PM PT.

Go to https://claude.ai/settings/usage, turn on extra usage and enable the promo from the notification afterwards.

I received €42, top up was not required and auto-reload is off.

Good question! Since this is an extension of input, I'm not sure if this is defined: https://developer.mozilla.org/en-US/docs/Web/API/HTMLInputEl....

On my desktop Chrome on Ubuntu, it seems to be persistent, but on my Android phone in Chrome, it loses the directory if I refresh.

I see this as a way to build apps with agentic flows where the original files don't need manipulation; instead, you create something new. Whether it's summarizing, answering questions, or generating new documents, you can use a local/internal LLM and feel relatively safe when tool calling is also restricted.

Related https://news.ycombinator.com/item?id=12098338

I like the perspective used to approach this. Additionally, the fact that major browsers can accept a folder as input is new to me and opens up some exciting possibilities.

Distributing pre-built images can help minimize this: https://hub.docker.com/r/nezhar/claude-container/tags.

I'm not sure how large the image becomes on the VM or if the distribution process is straightforward.

I also had the same idea when I built https://github.com/nezhar/claude-container.

What I was also curious about is what is actually sent and received by the agent, so I included this feature and created a CLI to make integration easier in a developer workflow.

Since I started doing this for other agents as well, I considered the idea of using a VM with Vagrant. However, I want the setup to remain minimal, so I still believe there is room for improvement.

The first rule covers it best: it is crucial that you know what you are doing in order to benefit from these tools.

Regarding the second and third rules, I like to work with two terminals. One is for my agent, isolated in a container, and the other one I use for `git`. This way, I can assess whether my changes are moving in the right direction.

The PLAN.md files have also proven to be a good approach. I'm still trying to figure out the best way to keep them aligned with the application without polluting the context, but they hold significant value for documentation purposes.

It's really hard to follow them; it feels like the Spider-Man meme.

https://nezhar.com/

This is a great example of why network restrictions on an application are not sufficient.

In my opinion, having a container is currently the best trade-off in terms of performance and maintainability of the setup.

Looks interesting. How does this compare to a container?

The isolation pattern is a good starting point.

I believe the detection pattern may not be the best choice in this situation, as a single miss could result in significant damage.

I built https://github.com/nezhar/claude-container for exactly this reason - it's easy to make mistakes with these agents even for technical users, especially in yolo mode.

Author here. This release adds two features I've found useful for understanding Claude Code behavior:

1. API Proxy - Transparently logs all interactions with the Anthropic API. Every request/response is captured without modifying Claude Code itself.

2. Datasette Integration - Lets you query and visualize the captured API data with SQL. Useful for tracking token usage, analyzing prompt patterns, or debugging unexpected behavior.

The container itself provides isolation from the host system while maintaining persistent credentials and workspace access via bind mounts.

Happy to answer questions about the implementation or use cases.

You still need to connect to Anthropic and obtain an authorization token.

The isolation here refers to the workspace. Since you run the CLI in a container, the process can only access what you have mapped inside. This is helpful if you want to avoid issues like this: https://hackaday.com/2025/07/23/vibe-coding-goes-wrong-as-ai...