Cli Agent Mcp

cli-agent-mcp

[](https://github.com/Bytars/cli-agent-mcp/actions/workflows/ci.yml) [](https://github.com/Bytars/cli-agent-mcp/actions/workflows/release.yml) [](https://github.com/Bytars/cli-agent-mcp/releases/latest) [](https://pkg.go.dev/github.com/Bytars/cli-agent-mcp) [](./LICENSE)

A single-binary **MCP (Model Context Protocol) stdio server** that lets an MCP client — such as **Claude Desktop** — drive a **local headless CLI coding agent** as a background worker, with **live progress streaming**.

Out of the box it drives **Claude Code** and **Cursor**, and any other CLI tool can be wired up with environment variables alone — no code required.

Instead of copy-pasting between a chat window and a terminal agent, the client delegates a task, watches the worker narrate its progress in real time, and gets the result inline — as if it had done the work itself.

Why run the agent locally?

The worker is spawned as a **child process of this server and inherits its environment**. That means it can reach whatever the host machine can reach:

• private networks and VPN routes,
• internal hosts via your SSH agent (1Password, ssh-agent, Pageant, …),
• cloud CLIs, kubeconfigs, and any credentials already set up on the box.

The MCP client never sees your keys or tokens — it just delegates a task to a worker that already has access.

MCP client (e.g. Claude Desktop)
        │  MCP over stdio
        ▼
  cli-agent-mcp                      ← this project
        │  spawn (headless)
        ▼
  claude / cursor-agent / your CLI   ← inherits env: VPN, S