Renting GPUs to run your own model
Mercury is a coding agent that runs in your terminal, and for version 1.0.0-beta.42 the thing I care about most is not depending on a subscription to run it. This release makes that real.
You've always been able to run Mercury on your own machine, with your own model weights. But serious models need a serious GPU, and that's been the catch: renting one, getting the model onto it, keeping it running. It works, but you spend more time on the rental than on your work.
So in this release, the model screen in Mercury (/llmhost) now searches public Hugging Face text-generation models by name to find the model you want. It comes with 78 ready-made setups for 32 models, with 45 new ones for the big open models like Reflection, Hy4, Gemma 4, and gpt-oss-120b. And before you rent anything, Mercury reads the model headers and config and refuses the ones that won't run, so you don't pay for a machine that dies on load.
Four rental providers are right there on every model's screen: Runpod, Vast.ai, Lambda, Hugging Face, cheapest first, with each recipe marked as measured, documented, community or unverified and each provider row marked with its own verification status. And one "Bank" screen where you download a model's weights to your own folder, verify every file, copy them somewhere safe, or mirror them onto storage you rent, so you don't download them again on every boot.
If a rental's engine dies, a download stalls, or a cluster fails, Mercury kills the machine, double-checks it's gone, and tells you what happened and whether you're still paying. I added this because I lost my own money on dead rentals before this release. If a running machine goes quiet, Mercury checks the engine first and says "here's how to restart it, here's how to stop the billing."
Why does this matter? I want the product to be straight with you about money. Not "GPU time included" when it's a credit that runs out. Not "estimated cost" when overages are on you. I want Mercury to rent only the machines that work, only when you use them, and I want it to mind your billing the way you would. This release gets closer.
The other big change is computer use. Now when Mercury clicks, types or waits, the result comes back as text and what has focus instead of a screenshot every time. One set of actions means one screenshot at the end, and window captures default to half size. And in a Mac terminal window, Mercury reads the text instead of the pixels. A coding agent shouldn't need photos to see what's in a terminal. Upshot: far fewer pixels going up to the model on each turn, so it's cheaper and faster.
A pile of smaller things made it in too: the debugger can read expressions and expand the result; Edit and Write catch new syntax errors in TypeScript, Python, Rust, Go and C with no language server; and the bundled Node runtime is updated to 24.22.1 (CVE-2026-111497). The CLI is BUSL-1.1, the SDKs are Apache-2.0.
What's next: measuring provider storage on real rentals, publishing your weights to your own private Hugging Face repo, and choosing where weights come from in the first place. But right now, /llmhost in Mercury is open for business. Rent an RTX PRO 6000 or an H200, point it at the model you want, and go.
You can see the full release notes here: https://github.com/Whq02/MercuryCLI/releases/tag/v1.0.0-beta.42