New stateful server tool: Shell! Any model on OpenRouter can now run commands in a hosted Linux container and use the new Files API to move files in and out. Available today in beta: `openrouter:shell`

Add the server tool to your request and the model decides when it needs a terminal. We support both the native OpenAI shell tool and Anthropic bash tool specs. Just add it to your tools array: { "type": "openrouter:shell", "parameters": { "engine": "openrouter" } }

The Files API means models aren't limited to what they can read/write natively. Upload any file type, and the model writes a script to process it. Ask for any file type, and it writes a script to produce it.

Pricing is designed to keep tracking simple by billing the computer entirely within the request. It's $0.0001 per active second, inclusive of file usage, with a 30-second minimum for cold container starts.

Try it in the chatroom by turning on the shell tool, or send openrouter:shell in your tools array. Read more at:

Love what you guys are building. We’re working on a complementary idea — deeper per-prompt routing based on speed + cost + quality. Excited to see how the multi-model space keeps evolving.

Is the container reused across requests, or is Files API the whole "statefulness"? 30s cold start is brutal for interactive loops if it's per-request.

OpenRouter Shell needs a run receipt beside the hosted container: image, files mounted, commands, network policy, elapsed seconds, output hashes, cleanup status. Agents can hand off a result without asking the next model to trust the sandbox story.

Ooo interesting

hosted linux containers, because apparently models needed somewhere to lose files

nani!?

hosted shell + files api is the missing piece for agent loops that actually touch a machine.

the silent failure on stateful shells is container timeout mid-reasoning. model assumes cwd is still intact, state already reset. no error, just output that looks like a logic bug for 20 minutes.

Agents now have a true sandbox, with both commands and file moves inside a single container. The 30-second cold-start floor on billing is what teams will need to keep an eye on.

1/ With beta access, devs can test file transfer pipelines at scale. 500+ models behind it.

Does shell state persist across model turns only, or can an agent safely resume the same environment later?

That's awesome! This new tool sounds incredibly useful. I'm excited to see