cosmicstack-labs

mercury-agent: Run a Permission-Aware AI Agent Across Channels

Mercury is a self-hosted AI agent for CLI, Telegram, and web use, with permission prompts, persistent memory, token budgets, and scheduled tasks. It suits people who want an always-on assistant they can configure and operate locally.

TypeScriptMITAI AgentsLLMCLI
View on GitHub Compare
Stars
3.2k
Forks
341
Stars, last 30 days
+602

Our take

Health score

60/100

How it is scored
Activity
83
Community
35
Issues
51
Pull requests
70

+602 stars in the last 30 days

  • Releases more often than 83% of the projects we track

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

4,035

Forks

410

Change since Nov 17, 2026: +6 stars·0 forks

Overview

Mercury is a TypeScript AI agent that connects to language-model providers and exposes the agent through a CLI, Telegram bot, and local web dashboard. It addresses the gap between conversational assistants and agents that can take actions by adding approval controls, token budgets, and persistent memory.

It is aimed at individuals who want a customizable assistant that can stay running, remember selected context, and work across channels. Its broad set of capabilities, including shell and file tools, scheduling, and task boards, makes it more involved to configure and supervise than a basic chat interface.

Key Features

  • Approval modes and folder-level read/write scoping for tools that access files, run commands, or interact with external systems.
  • SQLite-backed Second Brain memory with full-text search, automatic extraction, recall, and user controls.
  • CLI, Telegram, and local web dashboard interfaces, including streaming chat.
  • Daily token budgets with usage tracking and budget controls.
  • Background daemon with crash recovery and platform service integration for starting on boot.
  • Built-in filesystem, shell, Git, web, messaging, skills, and scheduling tools.
  • Kanban boards with task execution, dependencies, and per-card token tracking.
  • Provider fallback across configured hosted and local model options.

Use Cases

  • Individuals who want an assistant available through Telegram while Mercury runs as a background daemon.
  • Developers who want an agent to handle approved file, shell, and Git tasks from a CLI or workspace interface.
  • Users who need an assistant to retain preferences and project context across conversations on their own machine.
  • Teams or individuals who want to schedule recurring tasks or let an agent process cards on a persistent Kanban board.

Getting Started

For macOS or Linux, install the standalone binary with:

curl -fsSL https://mercuryagent.sh/install.sh | sh

The first run starts a setup wizard. See the README for other installation options and configuration details.

Alternatives

  • openinterpreter: Open Interpreter focuses on terminal-based coding and computer tasks, while Mercury offers a persistent assistant across CLI, Telegram, and web with scheduled tasks.
ProjectLanguageLicenseStarsStatus
mercury-agentTypeScriptMIT3.2kNot checked yet
openinterpreterRustApache-2.068.5kActive

Considerations

  • Using the npm package requires Node.js 20 or later. The project also offers standalone installers that do not require Node.js on the target machine.
  • Mercury needs credentials for a configured hosted model provider, unless you use its local Ollama option.
  • Permission controls reduce the chance of unintended actions, but users should still review what they approve and take care when granting broad access.
  • Community-contributed skills are described as unaudited, so review them before installation.
  • The project has 3.2k stars and 341 forks, but these counts alone do not establish suitability for a particular security-sensitive or production workload.

Found this useful?

Share it with someone who would like mercury-agent.

Comparisons