Mercury Agent
Soul-driven AI agent with permission-hardened tools, token budgets, and multi-channel access. Runs 24/7 from CLI or Telegram.
Last verified:
What is Mercury Agent?
Mercury Agent is a soul-driven, permission-aware AI agent that thinks, acts, and asks before executing tasks. It orchestrates multi-agent workflows with parallel sub-agents, intelligent loop detection via Mercury Autopilot, and 40+ built-in tools requiring zero configuration. The agent runs 24/7 as a daemon from your terminal or Telegram, featuring real-time progress tracking with live status cards that update in place as tools execute.
Key features include multi-agent orchestration with file locks preventing write conflicts, SQLite-backed persistent memory called "Second Brain" with 10 typed categories and conflict resolution, and a Markdown-based skill system with 126+ vetted skills across 23 categories. Mercury supports any LLM provider with automatic fallback (ChatGPT Web, GitHub Copilot, DeepSeek, OpenAI, Anthropic, Grok, Ollama Cloud/Local), daily token budgets with auto-concise behavior at 70%, and editable markdown "soul files" defining personality. It offers three access channels: CLI with ink-based TUI and workspace IDE mode, web dashboard at localhost:6174 with Kanban boards and memory graph visualization, and Telegram with inline keyboards and pinned status cards.
Mercury is designed for developers, technical users, and anyone wanting a persistent AI assistant that handles filesystem operations, shell commands, Git/GitHub PRs and issues, Spotify playback with DJ mode, web search, and scheduled tasks. The MIT-licensed TypeScript project aims to be the agent you set up once and leave running, with crash auto-restart, system service installation for auto-start on boot, and directory scoping with per-session memory.
Mercury Agent pricing
Pricing model: Freemium
Mercury Agent is completely free and open-source under the MIT license. There are no paid plans, subscription tiers, or usage fees. The agent itself costs nothing to install and use. Users only pay for their chosen LLM provider API costs (OpenAI, Anthropic, DeepSeek, etc.) or can use free options like ChatGPT Web with existing Plus/Pro subscription, GitHub Copilot subscription, or local Ollama models at zero cost. No credit card required.
Mercury Agent pros
- Multi-agent orchestration with parallel workers and file locks
- Intelligent loop detection via Mercury Autopilot analyzing parameter diversity and success rates
- Real-time progress tracking with single edited status card pinned on Telegram
- 40+ built-in tools requiring zero configuration
- Two permission modes: Ask Me for full control and Allow All with AI self-monitoring
- SQLite-backed Second Brain memory with 10 typed categories and conflict resolution
- 126+ vetted skills in registry across 23 categories with one-command install
- Multi-provider LLM support with automatic fallback and last-successful tracking
- Daily token budgets with color-coded stats and auto-concise at 70%
- Native Spotify integration with playback, DJ mode, and 14 tools
- Editable markdown soul files for customizable personality
- GitHub companion with PRs, issues, and co-authored commits
- Three access channels: CLI TUI, web dashboard, and Telegram
- 24/7 headless operation as daemon with crash auto-restart
- Workspace IDE mode with file explorer, git panel, and keyboard shortcuts
- Open source under MIT license
- Standalone binary install requiring no Node.js
- Safe command whitelist where reads never prompt
- Directory scoping with per-session memory
- Person tracking with alias resolution and relationship graph visualization
Mercury Agent cons
- Requires Node.js 20+ for npm install path
- Web dashboard only accessible locally at localhost:6174
- Telegram integration requires bot pairing setup
- Daily token budgets may limit long-running tasks without override
- Ask Me mode can be interruptive for power users
- No cloud-hosted version available
- Learning curve for skill system and frontmatter format
- Daemon management adds complexity compared to one-off agents
- Memory data stays fully local limiting multi-device sync
- Some features like Kanban boards require web dashboard enablement
Frequently asked questions about Mercury Agent
What is Mercury Agent?
Mercury Agent is a soul-driven AI agent that thinks, acts, and asks permission before executing. It orchestrates multi-agent workflows with 40+ built-in tools, runs 24/7 from CLI or Telegram, features intelligent loop detection via Mercury Autopilot, SQLite-backed Second Brain memory, and supports any LLM provider with automatic fallback.
How do I install Mercury Agent?
You can install via npm with 'npm i -g @cosmicstack/mercury-agent' (requires Node.js 20+), or use the standalone binary installer 'curl -fsSL https://mercuryagent.sh/install.sh | sh' on macOS/Linux which needs no Node.js. Windows users use 'irm https://mercuryagent.sh/install.ps1 | iex' in PowerShell.
How do I run Mercury continuously?
Use 'mercury up' which installs the service, starts the daemon, and ensures it keeps running with crash auto-restart. This runs Mercury 24/7 as a background process. Alternatively 'mercury start' starts as background daemon, and 'mercury service install' installs as system service for auto-start on boot.
What are Ask Me mode and Allow All mode?
Ask Me mode prompts for permission before every write, command, and scope change giving full control. Allow All mode auto-approves actions but includes AI self-monitoring that evaluates progress and stops stuck patterns. Safe commands (reads) never prompt in either mode.
What is Mercury Autopilot?
Mercury Autopilot is intelligent loop detection that analyzes parameter diversity and success rates. High diversity (>60%) and success (>70%) means productive work continues. Moderate repetition triggers AI self-check in Allow All or prompts in Ask Me. Low diversity with high failure automatically stops the execution path.
How does the Second Brain memory work?
Second Brain has conscious mind (active working memory for current reasoning) and subconscious mind (long-term persistent recall). It stores memories in 10 typed categories in SQLite, resolves conflicts by choosing higher-confidence memory, auto-consolidates hourly with profile summaries, and tracks people with alias resolution and relationship mapping.
Which LLM providers does Mercury support?
Mercury supports ChatGPT Web (OAuth, no API key), GitHub Copilot (OpenAI/Anthropic/Google models), DeepSeek (default), OpenAI (GPT-4o-mini, GPT-4o, o3), Anthropic (Claude Sonnet/Haiku/Opus), Grok via xAI, Ollama Cloud, and Ollama Local. It auto-fallbacks to last-successful provider and lets you switch models mid-session with '/models use'.
What are Mercury skills?
Skills are Markdown-based extensions teaching Mercury specific tasks. Each is a single SKILL.md file with frontmatter (name, description, intents, allowed-tools) and Markdown instructions. The registry has 126+ skills across 23 categories. Install via CLI 'mercury skills install', dashboard URL installer, or chat command.
How do I access the web dashboard?
Enable it with 'mercury doctor' or edit ~/.mercury/mercury.yaml setting 'web: enabled: true'. The dashboard runs at http://127.0.0.1:6174 when Mercury is running. Features include chat with SSE streaming, Kanban boards, Second Brain memory graph, workspace IDE, and management for providers, skills, schedules, and permissions.
Is Mercury Agent free?
Yes, Mercury Agent is completely free and open-source under the MIT license. There are no paid plans or subscription fees. You only pay for your LLM provider API costs, or use free options like ChatGPT Web with existing subscription, GitHub Copilot, or local Ollama models at zero cost.