deepseek

DeepSeek MCP Server exposes tools to interact with DeepSeek AI models (Chat and Reasoner) via MCP-compatible clients. It supports multi-turn sessions, model fallback with circuit breaker, function calling, thinking mode, and JSON output mode to provide structured data. Your MCP client can invoke the deepseek_chat tool for conversational interactions, with optional session_id to preserve context across requests and tools for function calling. You can also enumerate and manage sessions using the provided deepseek_sessions tool. The server is designed to integrate with Claude Code, Gemini CLI, and any MCP-compatible client, enabling cost tracking, multimodal input if enabled, and a rich set of prompts/templates for debugging and code-oriented use cases.

Prerequisites:

• Node.js 18+ installed on your system
• Access to a DeepSeek API key (get one at https://platform.deepseek.com)

Manual installation (global):

npm install -g @arikusi/deepseek-mcp-server

From source (recommended if you want to build locally):

# Clone the repository
git clone https://github.com/arikusi/deepseek-mcp-server.git
cd deepseek-mcp-server

# Install dependencies
npm install

# Build the project (if using TS sources)
npm run build

Usage is then configured via the MCP client configuration (see mcpConfig).

Tips and notes:

• Ensure you have a valid DEEPSEEK_API_KEY and set it in your MCP config or environment before running.
• You can enable multimodal input by setting ENABLE_MULTIMODAL=true in your environment when supported.
• The server supports session management via the deepseek_sessions tool; use session_id to maintain context across requests.
• When using function calling, you can supply up to 128 tool definitions via the tools parameter in the deepseek_chat call.
• If running locally behind a firewall or proxy, verify network access to https://platform.deepseek.com for API calls.
• For Claude Code or Gemini CLI integration, add the npx command to your project or global MCP configuration as shown in the Quick Start notes.
• For production deployments, consider wiring circuit breaker and caching settings to handle model fallback gracefully.