Notes
- Remote files over ssh, copy-on-write worktrees, project rule files and 1.0.11
Mercury CLI 1.0.11 lets the file tools work over ssh and starts isolated worktrees with copy-on-write copies of git-ignored dependencies.
- TensorFold in Mercury CLI and 1.0.6
Mercury CLI 1.0.6 adds first-class support for TensorFold across Apple silicon, Linux CUDA, LAN pairing, rented GPUs and Windows.
- Swarm chat in the main chat, captains, declared waits, and 1.0.3
Mercury CLI 1.0.3 brings crewmate board messages into the main chat, adds /general, /team and /whisper, captains, declared waits, and local-model fixes.
- Mercury CLI 1.0.0 is here
Your coding assistant for the terminal, now with a real desktop, documents, local images and video, and training.
- Mercury CLI goes open source
Mercury CLI 1.0.0-beta.44 is open source under the Apache License 2.0, connects 22 model providers and switches a chat to its rented GPU on its own.
- Rented GPUs you can trust, creative apps in your terminal
Mercury CLI 1.0.0-beta.43 pauses a failing rental instead of ending it, adds three GPU providers, a weights chooser, Craft extensions and usage history.
- Renting GPUs to run your own model
Mercury CLI 1.0.0-beta.42 searches Hugging Face models in /llmhost, banks and mirrors weights, rents GPUs across four providers and cuts screenshots.
- The Mercury CLI SDK, structured Samples, faster Computer use, and 1.0.0-beta.41
Mercury CLI 1.0.0-beta.41 ships mercury-cli-sdk 0.1.0 for TypeScript and Python, structured Sample documents, targeted Computer use, and the Video tool.
- Model hosting with /llmhost, GPU rentals, and /restart for live sessions
Run models on your computer or a GPU from Runpod, Vast.ai or Lambda in /llmhost, switch a live session with /restart, and browse MCP tools.
- Browsing, faster turns, and workspace safety floors
Persistent browser profiles, keyless APIs for WebFetch, early streaming tools, archive and SQLite reads and writes, and five safety floors.
- Testing Nemotron 3.5 Bolt locally, GPU-aware local setup, and full IDE tools for crewmates
Local models finish multi-step tool loops and size to discrete GPUs, crewmates get the full IDE toolset, and we test Nemotron 3.5 Bolt.
- Faster memory search on Linux and Windows, broader test debugging and clearer mid-turn replies
Memory search by meaning is 15x faster on Linux and Windows, test debugging adds Vitest, Jest, Go and Cargo, and mid-turn replies stay clear.
- Upgrading Mneme, keyless web search and cleaner page reads
Mneme ranks memory by relevance and past use, catches contradictions, web search needs no API keys, and WebFetch returns markdown directly.
- Consolidating the recent changes
A few new changes and fixes after the bigger updates, with live model discovery, more consistent Sovereign mode and stricter memory checks.
- A thread between sessions
Each session leaves a note the next one reads at start, the prompt is under half its old size, and here is how we tested both before shipping.
- Le Chonk, Nous Portal and a busy machine
Mistral and Nous Portal keys now work in Mercury, a slow or busy machine no longer loses a session, and the rest is in the release notes.
- Tools that tell you what's happened
This update changes which tools travel with a request, what their results show and when permission is needed.
- Long chats, calmer view. How .28 builds on reliability
Automatic compaction that recovers instead of giving up, counts kept to the active session, images on text-only models and a calmer chat view.
- Your rules, word for word: memory a coding agent cannot rewrite
Persistent memory that keeps your rules word for word and shows why each fact came up, with fixes to resume, retries and the watchdog.
- The Advisor, new providers, and more
New providers, per-chat advisor control and compaction fixes in beta.26, with plans to pair a local model with an advisor.
- Clearer communication between crewmates
More coherent crew communication, stronger file claims, better tool access and the fixes in beta.24.
- Qwen, Mercury, and future plans
Local model testing, the fixes in beta.23, and hopes for more powerful models running locally through Mercury.
- Mercury beta.22: the crew moves into the chat
I’ve been building beta.22 with a fleet of agents on different models, using Jev alongside them to help decide what needs a closer look.
- Mercury beta.21: better control of a running session
Beta.21 improves editing, adds crew pause and resume, rebuilds the model picker and fixes state updates behind the terminal UI.
- Say hello to Jev! But he can’t say hello back...?
Jev joins Mercury in 1.0.0-beta.20 as a second-opinion tool, alongside usage, sign-in and interface fixes.
- Testing beta.19 across two machines
Developing 1.0.0-beta.19 meant working through a browser failure and splitting tests across two machines without dropping the gates.
- Keeping Mercury's agents in the right checkout
Worktrees, paused agents and recovery in 1.0.0-beta.18, found while using Mercury to develop Mercury.
- Building Mercury inside Mercury
What developing Mercury in its own sessions exposed about search, editing and test isolation, through 1.0.0-beta.17.