Mercury Agent is a personal AI agent that asks before it acts and keeps a persistent memory of its user. It runs from a terminal UI, a web dashboard, or Telegram, and can run around the clock as a background daemon that restarts after crashes and starts on boot on macOS, Linux, and Windows. It ships as an npm package or as standalone binaries that need no Node.js.
Every action that touches files, the shell, the network, or git goes through a permission system: a blocklist of dangerous shell commands that never execute, folder-level read and write scoping, a pending-approval flow, and a per-session choice between asking and allowing all. A daily token budget is enforced, with responses made more concise past 70% of the budget and per-request overrides.
The Second Brain is a local SQLite memory with full-text search. After each conversation it extracts facts across ten memory types (identity, preference, goal, project, habit, decision, constraint, relationship, episode, reflection), injects the most relevant ones before each message, resolves conflicts by confidence or recency, consolidates hourly into profile summaries and reflections, and prunes stale entries. The agent's personality is defined in Markdown files the user owns.
The agent loop is built on the Vercel AI SDK, and providers are tried in order with automatic fallback: DeepSeek by default, then OpenAI, Anthropic, Grok, Ollama Cloud, and local Ollama, plus custom OpenAI-compatible endpoints. Skills follow the Agent Skills specification and install from a community registry of more than 126 skills.
Features
- Permission-hardened tools: a shell blocklist, folder-scoped access, and an approval flow with Ask Me or Allow All modes
- Second Brain memory: SQLite with full-text search, automatic fact extraction, recall, conflict resolution, consolidation, and pruning
- Token budgets: daily limits with automatic concise mode and per-request override
- Channels: an Ink terminal UI, a React web dashboard, and Telegram with admin and member roles
- Daemon mode: background service with crash recovery and auto-start on boot across macOS, Linux, and Windows
- Built-in tools: filesystem, shell, git, web fetch, messaging, skills, and scheduling
- Skills registry: search, view, install, and update community skills from the terminal, web dashboard, or Telegram
- Scheduler: recurring cron tasks and one-shot delayed tasks that reply on their originating channel
- Kanban boards: cards processed by the agent in dependency order with per-card token tracking
- Provider fallback: ordered providers with automatic failover to the next one
- Coding sub-agent: delegate coding tasks to a background sub-agent with plan and execute modes
