Repository navigation
☿ Mercury Agent v1.5.2
Watch your bots work live, approve things one at a time, and get complex jobs done properly. Chats run at the same time, memory keeps one copy of each thing, and text Mercury reads can't steer it.
Getting real work done
- Complex chat work follows Mercury Code's rules. Once the chat makes a plan:
- it acts instead of announcing;
- it checks what it made before calling it done;
- when a tool fails, it fixes the cause or asks, never quietly switching methods.
- Commands run on Mercury's own Node. Node tools like HyperFrames no longer fail on an older system Node.
- A video that never rendered isn't sent as finished. A video that's all black, or never moves, is caught before it goes out.
- Plans keep up. A forgotten last tick is completed. Unfinished plans show "paused" or "stopped" instead of a step stuck "in progress".
- Skills and services you name are the ones used, for the whole chat. "Use hyperframes" or "with ElevenLabs" sticks, and Mercury also reads skills installed for other agents (
~/.agents/skills). - Voice notes say only the words. Directions like "keep it natural" are never read aloud.
Mercury Desktop
- Bots, live. A Live tab for every bot and fleet shows each run as it happens: what it was asked, every step, its reply as it writes it, and how it ended. Message any bot from there, or reply to a finished run.
- "Needs you", one at a time. Approvals are a card above the message box, in plain words, as a carousel when there are several.
- A new chat that knows you. It opens with suggestions made from what you're working on, and the cursor is ready in the message box.
- Image batches come as one group, with a full-screen slider.
- Artifacts: delete for real, see the space they take, and clean up in bulk.
- Fewer false bot approvals. Read-only commands in a bot's own folders no longer ask, and bot faces always have colour.
Memory
- One copy of each thing. "Heer is my wife" in eleven wordings is one memory now. Copies already saved are folded once, after a backup.
- People: one person under several names becomes one entry, and stray words are checked before they count as people.
Security
- Untrusted content can't steer privileged actions. After reading files, pages or other people's messages, actions that schedule, save, send or install need your yes. Hidden instructions are flagged, and data can't leave through a web address.
- Strong models are asked only before data leaves. Small, local and unknown models keep every approval.
- Far fewer false alarms on documentation about AI.
Also
- Chats run at the same time, faster and cheaper requests, images, video, music and speech as media, and per-chat effort. The full list is in the CHANGELOG.
Install
- Desktop: mercuryagent.sh/download · Mac:
curl -fsSL https://mercuryagent.sh/install-desktop.sh | sh - Server / CLI:
curl -fsSL https://mercuryagent.sh/install.sh | sh· Windows:irm https://mercuryagent.sh/install.ps1 | iex· npm:npm i -g @cosmicstack/mercury-agent - Already installed: the Update button in Mercury Desktop, or
mercury upgrade