Overview
Mercury Agent is an open-source local AI agent framework (MIT license) developed by Cosmic Stack Labs, with version 1.0 officially released in April 2026. Its core design philosophy is to create a personal AI assistant that 'has a soul, memory, and boundaries'—defining personality through soul files, persisting long-term memory via SQLite, controlling permissions with 34 security interception rules, and supporting 7×24 uninterrupted operation through CLI and Telegram dual channels.
Key Features
- 31 Built-in Tools: Covering file system, Shell, Git, GitHub (PR/Issues), web scraping, scheduler, skill system, etc., ready to use out of the box
- Permission Hardening System: 34 security interception rules, command blacklist, folder-level read/write scope control, default 'ask first, then do' to prevent misoperation and malicious attacks
- Second Brain Memory: SQLite + FTS5 full-text search implements 10 memory types (identity/preferences/goals/projects/habits/decisions/constraints/relationships/events/reflections), automatically consolidated every hour
- Token Budget Management: Daily token cap enforced, auto-trim replies when exceeding 70% to prevent cost runaway from agent loop calls
- Soul Personality System: Define AI personality through four files: soul.md / persona.md / taste.md / heartbeat.md, ensuring strong behavioral consistency
- 7×24 Daemon: One-click background operation with 'mercury up', auto-restart on crash (exponential backoff), supports auto-start on macOS/Linux/Windows
Use Cases
- Personal programming assistant (Git operations, code writing, file management)
- GitHub collaboration companion (auto-create PRs, review code, manage Issues)
- Private AI assistant controlled remotely via Telegram
- Scheduled automation tasks (Cron scheduling backups, generating reports)
- Privacy-focused local knowledge management / second brain
Pros
- MIT open source, completely free, all data stored locally
- Most comprehensive permission security mechanism among similar tools
- Persistent memory allows AI to truly 'remember' user preferences and context
- Supports multiple LLM providers (DeepSeek/OpenAI/Anthropic/Ollama), with automatic fallback
- One-click launch with npx, configuration in 30 seconds
Pricing
Mercury Agent itself is completely free and open source (MIT license). Operating costs depend on the chosen AI provider: approximately $0.14 per million tokens with DeepSeek, or zero API fees with local models via Ollama.
Summary
Mercury Agent fills the gap in local AI agents regarding security and memory capabilities, making it suitable for developers who value privacy, need long-term operation, and want a personalized memory for their AI assistant. Compared to Open Interpreter and Claude Code, it is more comprehensive in permission control, token budgeting, and Telegram remote access.