Rented GPUs you can trust, creative apps in your terminal
The last release let you rent a GPU from inside Mercury and run your own model. This one is about trusting it with your money and your machine.
Nothing in this release terminates a rented machine on its own any more. On an engine exit, an out-of-memory error, a download stall, an SSH key refusal, five failed tunnel reconnects or a cluster fault, Mercury pauses the record with the machine still running, closes the local tunnel, and posts one notice naming the cause, the machine, the hourly rate still billing, and Resume or Terminate in /llmhost. A machine stays up until you stop it, or until 2 hours pass with no model request or with a pause nobody answered. The last release killed a dead rental on the first fault. That was too much, and it was my call to change it.
Three more providers joined the /llmhost screen: Hyperstack, Verda and DigitalOcean, with live stock and hourly prices for 32 models across 78 recipes. A Weights chooser on the rental confirmation lets you take the weights from Hugging Face, from a private repository of yours (every shard checked against the recipe's manifest before it downloads on the rented machine), from this computer's Bank, or from a Kept volume on Runpod or Vast.ai. Exactly what I said was next in the last release.
Three bundled creative apps made it into /extensions: FilmCraft 0.12.0, PhotoCraft 0.14.0 and LightCraft 0.13.0. Each one comes with a skill and three tools: craft_status to inspect the local install without downloading anything, craft_install which asks for confirmation with the version and size before pulling the verified app and CLI, and craft_connect to attach to your open control window, launch it, or start a separate headless session.
mercury usage --history reports your token counts, cache-hit rate, estimated USD and unpriced turns across saved local chats grouped by UTC day, model and project. No usage is sent anywhere; the summary is built and cached locally. DeepSeek turns are priced by request start time against peak and off-peak UTC hours.
/rewind got single-file restore, Undo the rewind, and a 500 MB file-history cap. Restore <file> puts back one file at a checkpoint and leaves the other files and the conversation untouched; Undo the rewind before your next prompt brings back the pre-rewind bytes; the 500 MB usage and limit show under State cleanup in /health.
The Debug tool runs adapter-requested processes as session-managed jobs without opening a terminal window, streams their output tagged [debuggee] into Debug's own, and ends the tree on disconnect. Every stop wait takes a timeoutMs, and aborting a call cancels its in-flight requests at once. Read can return a clipped line whole with lossless: true, and Edit, Write and ChangeSet refuse to touch a line that was clipped in Read.
Smaller things landed too: Private recipes, a file chooser under More options that validates and saves your own recipes file; every screen from /llmhost to /health saying what you are looking at, what it costs and what to do next; /lsp restart for a single server, ending its whole process tree; deeper GGUF checks before a rental, down to tensor types, shard counts and duplicate names; Windows .cmd and .bat files run through cmd.exe with their arguments passed verbatim; custom compat endpoints send prompt_cache_key; /mercury-docs bundles only git-tracked files. When a running rental's engine dies after setup, Mercury keeps the machine and the weights, reads the engine log, and tells you how much context it held against the recipe's budget and whether to Resume or Tune. And /clear on a model that can no longer be reached starts the fresh chat on the default model with a one-line notice instead of failing.
Why does this release matter? I want Mercury to mind your costs the way you would. Not "GPU time included" when it's a credit that runs out. Not promising savings it can't deliver. This release gets closer: nothing ends a machine behind your back, nothing bills you quietly, and when a rental fails you hear why and what it costs.
You can see the full release notes here: https://github.com/Whq02/MercuryCLI/releases/tag/v1.0.0-beta.43