MCP CLI is a feature-rich command-line interface for interacting with Model Context Protocol (MCP) servers, helping users complete tool calls and conversation management through an LLM. The default configuration uses Ollama and the gpt-oss reasoning model for local, privacy-focused operation without requiring an API key; it also supports cloud providers such as OpenAI, Anthropic, Azure, Google, and Groq, as well as custom providers.
Main capabilities
- Multiple operation modes: chat mode (streaming responses and automatic tool use), interactive mode (command-driven shell), command mode (suitable for scripts and pipelines), and direct commands.
- Advanced chat interface: streaming responses, visualization of the reasoning process, concurrent tool execution, intelligent Ctrl+C interruption, performance metrics, Markdown rendering, token usage tracking (/usage), multimodal attachments, session persistence, and conversation export (Markdown/JSON).
- Tool system: automatic discovery of server tools, provider-specific tool name adaptation, middleware (timeouts, retries, exponential backoff, caching, circuit breaking), multiple execution strategies (in-process, isolated subprocess, remote MCP), concurrent execution, and a complete audit trail.
- MCP Apps (SEP-1865): servers can provide interactive HTML UIs (charts, tables, maps, Markdown viewers) that run in a browser sandbox iframe, communicate in real time with the MCP server through a WebSocket bridge, and start automatically when tools annotated with _meta.ui are called.
- Execution plans (Tier 6): use /plan to create, inspect, and execute reproducible tool call graphs; supports model-driven planning (--plan-tools), parallel batch execution (Kahn BFS topological order), variable resolution (${var}, template strings), dry-run, checkpoints and resume, integration with budget/rate-limit guards, DAG ASCII visualization, and failure replanning.
- AI virtual memory (experimental, v0.16): --vm enables operating-system-style virtual memory management for conversation context, --vm-budget controls the token budget, --vm-mode selects passive/relaxed/strict, the /memory command visualizes the page table and eviction metrics, and supports multimodal page_fault and page export.
- Real-time browser dashboard (--dashboard): Agent terminal, activity stream, plan viewer, tool registry, configuration panel, file upload (including drag-and-drop and clipboard paste).
- Production hardening: automatic redaction of secrets in logs, structured rotating JSON logs (--log-file), per-server timeouts, thread-safe OAuth, and server health monitoring (/health).
- Performance and experience: O(1) tool lookup, LLM tool metadata caching, startup progress, automatic session saving, 8 built-in themes, cross-platform terminal management, and context-aware command completion.
Usage
- Prerequisites: Python 3.11 or later. For local operation, Ollama (ollama.ai) must be installed and the default reasoning model pulled: ollama pull gpt-oss. Cloud providers require setting the corresponding environment variables, such as OPENAI_API_KEY, ANTHROPIC_API_KEY, AZURE_OPENAI_API_KEY and AZURE_OPENAI_ENDPOINT, GEMINI_API_KEY, GROQ_API_KEY, etc.
- Enter chat, interactive, or command mode through the command line; commands such as /plan, /memory, /sessions, /export, /usage, /health, and /attach are available.
Applicable scenarios and limitations
- Suitable for users who need to interact with MCP servers, orchestrate tool calls, plan multi-step tasks, hold multimodal conversations, and run with local privacy as a priority.
- The official materials provide installation, configuration, command, and feature descriptions, but the specific installation commands, pricing, fees, and some platform compatibility details are not yet provided in the given materials; please refer to the official repository.
