Deepseek MCP Server

What arikusi/deepseek-mcp-server MCP server does

The arikusi/deepseek-mcp-server MCP server connects MCP-compatible clients to DeepSeek V4 models. It supports deepseek-v4-flash for faster, more economical requests and deepseek-v4-pro for higher capability; both provide a stated 1M-token context window. Older deepseek-chat and deepseek-reasoner names remain accepted as deprecated aliases.

The main chat tool accepts conversation messages and can generate regular or streaming responses. Requests may enable thinking mode, select reasoning effort, request JSON output, or provide a JSON Schema for response validation. Function calling is compatible with OpenAI-style tool definitions, with up to 128 tools per request.

A separate FIM tool fills text between a supplied prefix and optional suffix. It is intended for code completion and content infilling rather than conversational exchanges. The sessions tool lists, deletes, and clears stored conversation sessions.

How it works

The arikusi/deepseek-mcp-server MCP server communicates with an MCP client through stdio when run locally. The client invokes the exposed tools, while the server sends requests to DeepSeek using the configured API key. Conversation context can persist between calls when a session_id is used.

It also provides the deepseek://models, deepseek://config, and deepseek://usage MCP resources. These expose model information, server configuration, and usage statistics. Cost tracking includes cache-hit and cache-miss details. Model fallback and circuit-breaker behavior are available for handling model failures.

The project supports a hosted Streamable HTTP endpoint at https://deepseek-mcp.tahirl.com/mcp, so clients that support HTTP transport can connect without installing Node.js. The hosted option still requires the caller's DeepSeek API key.

Setup and configuration

Local installation requires Node.js 22 or newer and a DeepSeek API key from the DeepSeek platform. The documented package is @arikusi/deepseek-mcp-server, and a direct stdio configuration passes the key through DEEPSEEK_API_KEY.

A local client can launch the package with npx -y @arikusi/deepseek-mcp-server. The repository also documents installation from source by cloning the project, installing npm dependencies, and running its build script.

For remote use, configure the MCP URL and send an Authorization: Bearer header containing the DeepSeek key. HTTP mode is also available for self-hosted deployments through the TRANSPORT=http setting. Docker support is provided through the repository's multi-stage Dockerfile and health checks.

Tools and capabilities

• deepseek_chat for multi-turn chat, model selection, thinking mode, streaming, JSON output, schema validation, and function calling.
• deepseek_fim for prefix-and-suffix text infilling, including code completion.
• deepseek_sessions for listing, deleting, and clearing sessions.
• Three MCP resources for model, configuration, and usage data.
• Twelve prompt templates for tasks such as debugging, code review, and function calling.
• Optional multimodal content-part support for text and image input when enabled with ENABLE_MULTIMODAL=true.

Limitations and notes

DeepSeek API usage requires the user's own API key and may incur provider charges; the MCP package itself is MIT-licensed. FIM uses DeepSeek's Beta endpoint and operates in non-thinking mode. Thinking mode ignores the temperature setting. Schema validation can perform a bounded number of repair retries controlled by RESPONSE_SCHEMA_MAX_RETRIES.

The documented model names are the V4 names, while the older aliases are retained for compatibility. The server's listed client compatibility includes Claude Code, Gemini CLI, Cursor, Windsurf, and other MCP-compatible clients; support in a particular client depends on its MCP transport configuration.