Mercury
🔥 This release made it to our blog
Mercury: The Open-Source AI Agent That Thinks, Acts, and Asks Permission
Mercury is an open-source AI agent built around three principles: it thinks with intelligent loop detection and 25-step agentic loops, acts with 40+ built-in tools across filesystem, shell, GitHub, and Spotify, and asks — ensuring nothing happens without your explicit permission or AI self-monitoring.
### TL;DR
Mercury is an open-source AI agent (MIT) that thinks, acts, and asks permission. It features multi-agent orchestration with parallel sub-agents, intelligent loop detection via Autopilot, 40+ built-in tools (filesystem, shell, git, GitHub, Spotify), a Second Brain memory system, and runs 24/7 headless from your terminal or Telegram with full permission controls.
Key Insights & Metrics
Key Features
- Multi-agent orchestration with parallel sub-agents — file locks prevent conflicts, resource-aware concurrency, non-blocking chat while agents run; Mercury Autopilot detects stuck loops by analyzing parameter diversity (>60%) and success rate (>70%) and halts unproductive paths automatically
- 40+ built-in tools with zero configuration — filesystem, shell, git, GitHub PRs/issues, Spotify (14 tools + DJ mode), Markdown skill system with scheduling; real-time progress via single live-edited status card on CLI and Telegram with token stats and budget tracking
- Second Brain memory with conscious/subconscious layers — active working memory + persistent long-term recall in local SQLite; person tracking with alias resolution and graph visualization; three interfaces: CLI (TUI), Web Dashboard (React SPA at localhost:6174), and Telegram with inline keyboard permissions
Discussion
Sign in to leave a review
Reviews
No reviews yet. Be the first to review!