AgentMCP — Let AI agents be really autonomous without breaking your machine

> Give your AI agent true autonomy — without giving it the keys to your machine.

Installation

Windows

powershell -ExecutionPolicy Bypass -Command "irm https://raw.githubusercontent.com/kokojkj/agentmcp/main/scripts/install.ps1 | iex"

Linux

curl -fsSL https://raw.githubusercontent.com/kokojkj/agentmcp/main/scripts/install.sh -o install.sh && bash install.sh

Installation takes a while, so grab a cup of coffee ☕ — once it's done, the installer will guide you through connecting AgentMCP to Claude and ChatGPT.

Related MCP server: mcp-agent-tools

Why AgentMCP?

Working with AI agents today means dealing with frustrating trade-offs:

1. Permission fatigue — Agents constantly stop to ask for permission, killing your flow. The alternative, "always allow," hands the agent full access to do whatever it wants — whether that's finishing your project or wrecking your machine.

2. Rigid tool setup — Agents only know the tools and MCP servers you configured before the task starts. If the agent needs a new tool mid-task, it fails — and you have to step in and restart the whole task from scratch.

3. No safety net — An agent can modify or delete your files at any time, with no way to restore them and no file history showing what it changed.

4. No parallelism — Most agents don't support async command and MCP execution. A complex task with many commands runs them one by one, leaving the agent stuck waiting instead of firing them all in parallel.

AgentMCP fixes all four — enabling truly autonomous agents that do real work, without breaking your machine or slowing down your workflow.

Want the full breakdown of how it solves each problem? Read the deep dive - still WIP.

Current Status

• ⚠️ Alpha / work in progress — expect rough edges, occasional incorrect behavior, and some security issues.

• 🚧 Not production-ready yet — contributions are very welcome to help us get there.

Roadmap

• Minimize the codebase and improve efficiency

• Complete HTTP mode (currently incomplete)

• Resolve overlapping tools and make tool definitions more token-efficient

• Fix logic pitfalls in MCP "lifecycle" control (may be removed entirely)

Contributing

I am looking for contributors and Co-maintainers!

If you think this project is worth your time, we'd love your help making it better for everyone. To get started, read the contribution guide.