llm-openrouter 0.7 turns on server-side tools, fetch, and web search for models

Server-side tools are no longer just in the cloud

The new version of the llm-openrouter plugin (0.7) for the LLM CLI tool brings compatibility with version 0.32. This means that models available via OpenRouter can now use the implementation of the Responses API. The most important novelty, however, are three new server-side tools that the plugin makes available: Shell, WebFetch, and WebSearch.

CLI agents get new eyes and hands

For developers working with the CLI, this means that the agents they interact with are no longer locked in a text box. If you call a model and allow it, for example, WebSearch (via the -T WebSearch parameter), the model can reach out to the internet on its own, find current information, and work with it. Shell, in turn, allows it to run commands directly in the terminal. It's a shift from passive autocomplete to an agent that has access to the outside world and the system.

The risk of hands that are too free

Every authorization of server tools carries a security risk. If a model gets its hands on Shell and at the same time sucks data from the internet, the risk of prompt injection or inappropriate behavior grows exponentially. It is not difficult to imagine a situation where a model executes an unwanted command based on malicious input from the web.

What will be the limits

The key metric will be how smaller models handle these possibilities and whether developers can set boundaries for running Shell commands. If the tool catches on in normal workflows, it will show how ready models are for autonomous operations with access to the server.

Lilith's verdict

Giving a model access to the web via CLI is fine, but handing it a Shell is like handing a toddler a live grenade. Let's see who accidentally wipes their server first.