Mercury
A permission-hardened AI agent with token budgets and persistent memory that runs continuously from the command line, Telegram or the web, asking before it acts.
Mercury is a soul-driven AI agent with permission-hardened tools, token budgets and multi-channel access, designed to run continuously from the command line, Telegram or a web interface. It ships 31 built-in tools, Kanban boards and extensible skills, remembering what matters through a SQLite-backed Second Brain with ten memory types, automatic extraction, conflict resolution and consolidation.
Its permission system blocklists dangerous shell commands, scopes read and write access by folder, and requires pending approval for actions unless the session is set to allow everything. Its personality is defined entirely by markdown files the user owns, rather than a fixed persona, and a daily token budget is enforced with automatic conciseness above 70% usage and a command to check, reset or override it. It runs as a background daemon that restarts on crash, starts on boot, and supports cron scheduling and heartbeat monitoring.
It installs with a one-line script or through npm, and is written in TypeScript under the MIT license.