Soul-driven AI agent with permission-hardened tools, token budgets, and multi-channel access. Runs 24/7 from CLI, Telegram or More.
{
"schema": 1,
"issuedAt": "2026-11-12T14:28:06Z",
"server": {
"channels": {
"stable": {
"version": "1.5.2",
"date": "2026-11-12",
"notes": "## ☿ Mercury Agent v1.5.2\n\nWatch 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.\n\n### Getting real work done\n- **Complex chat work follows Mercury Code's rules.** Once the chat makes a plan:\n - it acts instead of announcing;\n - it checks what it made before calling it done;\n - when a tool fails, it fixes the cause or asks, never quietly switching methods.\n- **Commands run on Mercury's own Node.** Node tools like HyperFrames no longer fail on an older system Node.\n- **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.\n- **Plans keep up.** A forgotten last tick is completed. Unfinished plans show \"paused\" or \"stopped\" instead of a step stuck \"in progress\".\n- **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`).\n- **Voice notes say only the words.** Directions like \"keep it natural\" are never read aloud.\n\n### Mercury Desktop\n- **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.\n- **\"Needs you\", one at a time.** Approvals are a card above the message box, in plain words, as a carousel when there are several.\n- **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.\n- **Image batches** come as one group, with a full-screen slider.\n- **Artifacts:** delete for real, see the space they take, and clean up in bulk.\n- **Fewer false bot approvals.** Read-only commands in a bot's own folders no longer ask, and bot faces always have colour.\n\n### Memory\n- **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.\n- **People:** one person under several names becomes one entry, and stray words are checked before they count as people.\n\n### Security\n- **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.\n- **Strong models are asked only before data leaves.** Small, local and unknown models keep every approval.\n- Far fewer false alarms on documentation about AI.\n\n### Also\n- 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.\n\n### Install\n- **Desktop:** [mercuryagent.sh/download](https://mercuryagent.sh/download) · Mac: `curl -fsSL https://mercuryagent.sh/install-desktop.sh | sh`\n- **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`\n- **Already installed:** the Update button in Mercury Desktop, or `mercury upgrade`",
"notesUrl": "https://github.com/cosmicstack-labs/mercury-agent/releases/tag/v1.5.2"
},
"dev": null
},
"minimum": null
},
"desktop": {
"channels": {
"stable": {
"version": "1.5.2",
"date": "2026-11-12",
"notes": "## ☿ Mercury Agent v1.5.2\n\nWatch 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.\n\n### Getting real work done\n- **Complex chat work follows Mercury Code's rules.** Once the chat makes a plan:\n - it acts instead of announcing;\n - it checks what it made before calling it done;\n - when a tool fails, it fixes the cause or asks, never quietly switching methods.\n- **Commands run on Mercury's own Node.** Node tools like HyperFrames no longer fail on an older system Node.\n- **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.\n- **Plans keep up.** A forgotten last tick is completed. Unfinished plans show \"paused\" or \"stopped\" instead of a step stuck \"in progress\".\n- **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`).\n- **Voice notes say only the words.** Directions like \"keep it natural\" are never read aloud.\n\n### Mercury Desktop\n- **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.\n- **\"Needs you\", one at a time.** Approvals are a card above the message box, in plain words, as a carousel when there are several.\n- **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.\n- **Image batches** come as one group, with a full-screen slider.\n- **Artifacts:** delete for real, see the space they take, and clean up in bulk.\n- **Fewer false bot approvals.** Read-only commands in a bot's own folders no longer ask, and bot faces always have colour.\n\n### Memory\n- **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.\n- **People:** one person under several names becomes one entry, and stray words are checked before they count as people.\n\n### Security\n- **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.\n- **Strong models are asked only before data leaves.** Small, local and unknown models keep every approval.\n- Far fewer false alarms on documentation about AI.\n\n### Also\n- 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.\n\n### Install\n- **Desktop:** [mercuryagent.sh/download](https://mercuryagent.sh/download) · Mac: `curl -fsSL https://mercuryagent.sh/install-desktop.sh | sh`\n- **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`\n- **Already installed:** the Update button in Mercury Desktop, or `mercury upgrade`",
"notesUrl": "https://github.com/cosmicstack-labs/mercury-agent/releases/tag/v1.5.2"
}
},
"minimum": null
},
"messages": []
}