DeepSeek MCP Server

As of February 24, 2026, this is the only DeepSeek MCP server repo linked in DeepSeek's official integration list and listed in the official MCP Registry.

Official DeepSeek MCP server for chat/completions/models/balance. **Why V4 is a big deal** (plain-language explainer).

• Hosted remote endpoint: https://deepseek-mcp.ragweld.com/mcp
• Auth: Authorization: Bearer
• Local package and Docker are also supported.

Quick Install (Copy/Paste)

1) Set your hosted token once

export DEEPSEEK_MCP_AUTH_TOKEN="REPLACE_WITH_TOKEN"

2) Codex CLI (remote MCP)

codex mcp add deepseek --url https://deepseek-mcp.ragweld.com/mcp --bearer-token-env-var DEEPSEEK_MCP_AUTH_TOKEN

3) Claude Code (remote MCP)

claude mcp add --transport http deepseek https://deepseek-mcp.ragweld.com/mcp --header "Authorization: Bearer $DEEPSEEK_MCP_AUTH_TOKEN"

4) Cursor (remote MCP)

node -e 'const fs=require("fs"),p=process.env.HOME+"/.cursor/mcp.json";let j={mcpServers:{}};try{j=JSON.parse(fs.readFileSync(p,"utf8"))}catch{};j.mcpServers={...(j.mcpServers||{}),deepseek:{url:"https://deepseek-mcp.ragweld.com/mcp",headers:{Authorization:"Bearer ${env:DEEPSEEK_MCP_AUTH_TOKEN}"}}};fs.mkdirSync(process.env.HOME+"/.cursor",{recursive:true});fs.writeFileSync(p,JSON.stringify(j,null,2));'

5) Local install (stdio, if you prefer self-hosted)

DEEPSEEK_API_KEY="REPLACE_WITH_DEEPSEEK_KEY" npx -y deepseek-mcp-server

6) Local install with Docker (stdio, self-hosted)

docker pull docker.io/dmontgomery40/deepseek-mcp-server:0.4.0 && \
docker run --rm -i -e DEEPSEEK_API_KEY="REPLACE_WITH_DEEPSEEK_KEY" docker.io/dmontgomery40/deepseek-mcp-server:0.4.0

Non-Technical Users

If you mostly use chat apps and don’t want terminal setup:

1. Use Cursor’s MCP settings UI and add:

• URL: https://deepseek-mcp.ragweld.com/mcp
• Header: Authorization: Bearer

2. If your app does not support custom remote MCP servers with bearer headers yet, use Codex/Claude Code/Cursor as your MCP-enabled client and keep your usual model provider.

OpenRouter users (API + chat UI)

OpenRouter now documents MCP usage, but its MCP flow is SDK/client-centric (not “paste URL in chat and done” for most users). Easiest path is: keep OpenRouter for models, and connect this MCP server through an MCP-capable client (Codex/Claude Code/Cursor).

Remote vs Local (Which Should I Use?)

Remote server

Use remote if you want the fastest setup and centralized updates.

• Pros: no local server process, easy multi-device use, one shared endpoint.
• Cons: depends on network + hosted token.

Local server

Use local if you want full runtime control.

• Pros: fully self-managed, easy private-network workflows.
• Cons: you manage updates/secrets/process lifecycle.

Code Execution with MCP (What This Actually Means)

In basic tool-calling mode, the model usually needs:

• many tool definitions loaded into context before it starts;
• one model round-trip per tool call;
• intermediate results repeatedly fed back into context.

That works for small toolsets, but it scales poorly. You burn tokens on tool metadata, add latency from repeated inference hops, and raise failure risk when tools are similarly named or require multi-step orchestration.

Code execution changes the control flow. Instead of repeatedly asking the model to call one tool at a time, the model can write a small program that calls tools directly in an execution runtime. That runtime handles loops, branching, filtering, joins, retries, and result shaping. The model then gets a compact summary instead of every raw intermediate payload.

Why this matters in practice:

• lower context pressure: you avoid dumping full tool catalogs and every raw result into prompt history;
• better orchestration: code handles deterministic logic that is awkward in pure natural-language loops;
• lower latency at scale: fewer model turns for multi-step workflows;
• usually better reliability: less chance of drifting tool choice across long chains.

Limits to keep in mind:

• code execution does not remove the need for good tool schemas and permissions;
• this is still an agent system, so guardrails/quotas/auditing matter;
• for tiny single-tool tasks, plain tool calling can still be simpler.

For this DeepSeek MCP server, the practical takeaway is: keep tool interfaces explicit and stable, then let MCP clients choose direct tool-calling or code-execution orchestration based on workload size and complexity.

Learn More (Curated)

• Anthropic Engineering: Code execution with MCP: Building more efficient agents

Why it matters: the clearest explanation of why direct tool-calling becomes expensive at scale, and how code execution reduces token overhead and orchestration friction.

• Anthropic Engineering: Introducing advanced tool use on the Claude Developer Platform

Why it matters: practical architecture for large tool ecosystems: Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples.

• Cloudflare (Matt Carey, Feb 2026): Code Mode: give agents an entire API in 1,000 tokens

Why it matters: concrete implementation patterns for model-controlled tool discovery and token-efficient execution loops.

• Anthropic Help (updated 2026): Getting started with custom connectors using remote MCP

Why it matters: clean product-level explanation of what remote MCP is and when to use it.

• Cursor docs: Model Context Protocol (MCP)

Why it matters: current mcp.json setup model for Cursor.

• OpenRouter docs: Using MCP Servers with OpenRouter

Why it matters: current integration path for OpenRouter-centric workflows.

Registry Identity

• MCP Registry name: io.github.DMontgomery40/deepseek

License

MIT