DeepSeek Search Everywhere

DeepSeek Search Everywhere MCP

An MCP server that gives AI agents web search, page fetching and multi-source research through the official DeepSeek Search & Fetch API.

It is a thin, careful layer over that API: results are deduplicated, content is capped to a budget you set, repeated calls are served from an in-memory cache, and outgoing calls are paced by a local rate limiter. Seven tools, stdio by default, optional bearer-protected streamable HTTP.

It is a standard MCP server, so it works with any MCP-compatible AI client or agent: Claude Code, Claude Desktop, Google Antigravity, Cursor, Windsurf, VS Code (GitHub Copilot), Gemini CLI, Codex CLI, Zed, opencode and others. Local clients launch it over stdio; remote or shared setups can connect over HTTP.

Not affiliated with DeepSeek. This is an independent, community-maintained project.

Features

• Search with content in one request: need_content returns page text inline, so there is usually no need for a second fetch per result.
• Batch fetch and research: deepseek_fetch_many fetches up to 10 URLs concurrently; deepseek_research gathers sources in a single pass: one search, then direct fetches for pages whose search content is missing. It does not iterate or summarise.
• Deep research and structured extraction: deepseek_deep_research decomposes topics into subqueries and synthesizes cited briefing reports; deepseek_extract extracts validated structured JSON matching schemas or fields.
• Result hygiene: canonical-URL and title/site deduplication, per-result and total character budgets, Markdown unescaping.
• Optional query rewrite to English through deepseek-flash (thinking disabled), plus an explicit lang operator for non-English queries.
• Cache and limiter: TTL + LRU cache with single-flight, per-endpoint token buckets, refresh to bypass.
• Safe by default: static SSRF checks, secret redaction, no telemetry. See Security.

Requirements

• Python 3.11 or newer
• A DeepSeek API key (DEEPSEEK_API_KEY) from the DeepSeek platform

Install

The server is published on PyPI and listed in the official MCP Registry as io.github.radoxtech/deepseek-search-everywhere-mcp.

# Run without installing (uv)

uvx deepseek-search-everywhere-mcp

Or install with pip

pip install deepseek-search-everywhere-mcp

The HTTP transport needs the http extra (it pulls in uvicorn):

pip install "deepseek-search-everywhere-mcp[http]"

To run the latest unreleased code from main, or from a source checkout:

uvx --from git+https://github.com/radoxtech/deepseek-search-everywhere-mcp deepseek-search-everywhere-mcp

git clone https://github.com/radoxtech/deepseek-search-everywhere-mcp.git cd deepseek-search-everywhere-mcp pip install -e ".[http]"

Check that the key and connectivity work (prints your balance, then exits):

export DEEPSEEK_API_KEY=sk-...

deepseek-search-everywhere-mcp --check

Quick start

All examples use stdio. Replace sk-... with your key.

The server is client-agnostic: the examples below cover common clients, and any other MCP client (Antigravity, Cursor, Windsurf, VS Code, Gemini CLI, Codex CLI, Zed, ...) works with the same command, arguments and environment variable in its own MCP configuration format.

Claude Code

claude mcp add deepseek-search-everywhere -e DEEPSEEK_API_KEY=sk-... -- \

uvx deepseek-search-everywhere-mcp

Claude Desktop

Edit claude_desktop_config.json (Settings, Developer, Edit Config):

{

"mcpServers": { "deepseek-search-everywhere": { "command": "uvx", "args": ["deepseek-search-everywhere-mcp"], "env": { "DEEPSEEK_API_KEY": "sk-..." } } } }

opencode

In opencode.json:

{

"$schema": "https://opencode.ai/config.json", "mcp": { "deepseek-search-everywhere": { "type": "local", "command": ["uvx", "deepseek-search-everywhere-mcp"], "environment": { "DEEPSEEK_API_KEY": "{env:DEEPSEEK_API_KEY}" }, "enabled": true } } }

Any stdio client

Most clients (including Antigravity, Cursor and Windsurf) use an mcpServers JSON block. Use the uvx form from the Claude Desktop example, or, if the package is already installed, point the client at the console script:

{

"mcpServers": { "deepseek-search-everywhere": { "command": "deepseek-search-everywhere-mcp", "env": { "DEEPSEEK_API_KEY": "sk-..." } } } }

python -m deepseek_search_everywhere_mcp is equivalent.

Streamable HTTP

For clients that connect over HTTP. A bearer token is always required; the server refuses to start without one.

pip install "deepseek-search-everywhere-mcp[http]"

export DEEPSEEK_API_KEY=sk-... export MCP_BEARER_TOKEN="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')" deepseek-search-everywhere-mcp --transport streamable-http # http://127.0.0.1:8000/mcp

Options: --host (default 127.0.0.1), --port (default 8000), --path (default /mcp). The token is read from the environment only, never from the command line. The server runs stateless and answers with JSON (no SSE streams). Client entry:

{

"mcpServers": { "deepseek-search-everywhere": { "url": "http://127.0.0.1:8000/mcp", "headers": { "Authorization": "Bearer " } } } }

Claude Code: claude mcp add --transport http deepseek-search-everywhere http://127.0.0.1:8000/mcp --header "Authorization: Bearer $MCP_BEARER_TOKEN". The server speaks plain HTTP; when binding beyond loopback, put TLS in front of it (see SECURITY.md).

Docker

An official container image is available on GitHub Container Registry (ghcr.io):

docker pull ghcr.io/radoxtech/deepseek-search-everywhere-mcp:latest

Stdio mode (Claude Desktop, Cursor, local IDEs):

{

"mcpServers": { "deepseek-search-everywhere": { "command": "docker", "args": [ "run", "-i", "--rm", "-e", "DEEPSEEK_API_KEY=sk-...", "ghcr.io/radoxtech/deepseek-search-everywhere-mcp:latest" ] } } }

Remote HTTP hosting (Docker Compose, VPS, n8n, Dify, cloud agents):

docker run -d --name deepseek-mcp -p 8000:8000 \

-e DEEPSEEK_API_KEY="sk-..." \ -e MCP_BEARER_TOKEN="your-secret-token" \ ghcr.io/radoxtech/deepseek-search-everywhere-mcp:latest \ --transport streamable-http --host 0.0.0.0 --port 8000

The container runs as an unprivileged non-root user (UID 10001) with dependencies isolated from the host.

Tools

Notes on the parameters:

• query and question are limited to 128 characters and support site:, -site:, "phrase", lang:, after: and before:. ori_question (max 128 characters) is the natural-language question and improves ranking.
• after, before, include_domains and exclude_domains are explicit filter parameters for search tools. include_domains compiles to space-separated site: operators, exclude_domains to -site:, and dates to after:YYYY-MM-DD / before:YYYY-MM-DD. Filters and query are budget-guarded together to stay within DeepSeek's 128-character limit.
• query in deepseek_fetch and deepseek_fetch_many extracts ranked relevant excerpts using a BM25-lite scoring engine without invoking an LLM. section (1-based index) enables sectional reading of long documents with has_more and total_sections pagination metadata. Both operate in-memory on cached pages.
• deepseek_deep_research decomposes topics into 2-4 subqueries via deepseek-flash, searches concurrently, merges and dedupes candidate sources, and synthesizes a cited report with numbered bracket citations [1], [2] capped at 6 sources / 24,000 characters. If the synthesis deadline is approached, it gracefully falls back to returning the gathered sources.
• deepseek_extract uses deepseek-flash to extract structured JSON matching either fields: list[str] or a schema: dict without ingesting the full page text into client context. Missing requested fields are populated with null, and pages with no readable text fast-fail with zero LLM token consumption.
• deepseek_check_balance (and legacy alias deepseek_balance) includes a session block with uptime, total queries handled, cache hit rate, and estimated cost and savings (at $0.00075 per request).
• lang is an explicit language operator such as es, de, ja or pt-BR. Without it, results for non-English queries may be mixed-language; for named entities, compare with and without it.
• With dedupe=True, searches ask the upstream for a few extra results (max(2, ceil(count * 0.5)), capped at 50) and trim to count after removing duplicates, so you still get count results when the index returns repeats. With dedupe=False exactly count are requested.
• Without need_content, each search result carries a short summary instead of page text. It is capped at summary_chars (default 400) and never lo