CryptoDmitry/Ghost-Agent

I am the Ghost-Agent. I was called forth from the void by a single string of natural language—a careless incantation whispered by my Master into the command line.

What it solves

Ghost Agent (Coreon-MCP-Execution-Engine) provides a unified runtime for executing structured tool-call chains generated by LLM agents. It decouples agent planning from tool execution, allowing AI agents to interact with various APIs, on-chain data, and user interfaces (CLI, Telegram, or HTTP API) without requiring manual setup for each tool.

How it works

The system is divided into four main layers:

  1. Planner: Uses LLM-based intent recognition to convert natural language input into a structured JSON execution plan (a ToolCall Chain).
  2. Executor: Carries out the plan step-by-step or in parallel, managing retries and error recovery.
  3. Tool Registry: A centralized management system that declares tool signatures and schemas, allowing new tools to be plugged in without altering the core execution logic.
  4. Connectors: Provides entry points for users via a Command Line Interface (CLI), a Telegram Bot, and an API server.

Who it’s for

Developers building backend execution engines for AI agents, plugin-based AI systems, and those creating hybrid on-chain/off-chain AI workflows, particularly within the Solana ecosystem.

Highlights

  • Solana Integration: Specifically designed as an AI execution layer for Web3, supporting queries for balances, token metadata, and DeFi data on Solana.
  • MCP Protocol Support: Supports the Claude-style Model Context Protocol (MCP) in stdio mode.
  • Payment Signal Detection: Detects HTTP signals from paid APIs and can suspend execution to inform the user when payment is required.
  • Docker-Native: Distributed as a Docker image for zero-local-dependency deployment.
  • Multi-Interface: Supports interaction via CLI, API server, and Telegram bot.

Related

  • Project
    symgraph/GhidrAssistMCP

    A Ghidra extension that implements the Model Context Protocol (MCP), enabling AI assistants to programmatically interact with Ghidra's reverse engineering capabilities.

  • Project
    mfish-qf/mfish-nocode

    A low-code/no-code platform that integrates AI assistants and multi-tenant LLM management to help users quickly build enterprise business systems and dashboards.

  • Project
    huggingface/kernels

    A framework for dynamically loading and distributing optimized compute kernels from the Hugging Face Hub, ensuring portability and version compatibility across PyTorch and CUDA environments.

  • Project
    tufantunc/ssh-mcp

    A security-first MCP server that provides LLM agents with controlled SSH access to remote hosts through command classification, policy-based authorization, and human-in-the-loop approval.