Frontier Now
Search

Gateways

MCP CLI: connect to and debug any MCP server from the terminal

An MCP command-line client maintained under IBM. Call server tools by chat, interactive shell or scripts, with a local Ollama model by default and no API key.

Author
Frontier Now Editorial
Published

Project and installation docs

View project

https://github.com/IBM/mcp-cli

You’ve written an MCP server and want to see how a model actually calls it, which usually means editing Claude Desktop’s config, restarting and poking around. MCP CLI moves that loop into the terminal. It reads a server_config.json, connects to one or more servers, and lets you chat with a model that calls tools on its own, call a tool directly without a model, or pipe calls through shell scripts.

What it does

  • Four modes: chat mode lets the model pick tools; interactive mode is a command shell; command mode suits scripts and pipelines, e.g. mcp-cli cmd --server sqlite --tool list_tables --output tables.json; and single commands like tools, ping and resources.
  • Any model: defaults to Ollama’s gpt-oss running locally with no key, and supports OpenAI, Anthropic, Azure OpenAI, Gemini, Groq, watsonx and others.
  • Debugging aids: /usage tracks tokens, sessions can be saved and exported to Markdown or JSON, and --dashboard opens a browser view of the tool-call stream.
  • Newer protocol features: MCP Apps support opens tools with _meta.ui annotations as sandboxed browser UIs, and experimental execution plans turn multi-step tool calls into replayable graphs.

Who it’s for

  • Engineers building MCP servers who want to check tool descriptions and real calls quickly from a terminal.
  • People pairing local models with MCP tools, or wiring tool calls into batch scripts.

Setup

Needs Python and uv; for the default model, install Ollama and run ollama pull gpt-oss. List servers in server_config.json in the working directory (same mcpServers format), then:

uvx mcp-cli --help
mcp-cli --server sqlite

Our take

About 2,000 stars as of 2026-10-06. It’s an MCP client rather than a server for agents, and it’s here because debugging and scripting MCP tools is a real need that desktop chat apps handle clumsily. Compared with the official MCP Inspector, it adds a real model in the loop, which exposes vague tool descriptions. The downside is feature sprawl (virtual memory, execution plans, a dashboard) on top of the author’s family of chuk libraries, which can be disorienting; start with just the cmd and tools commands. Last push was in August 2026. Licensed Apache-2.0.