AI导哈

Serper搜索MCP工具包

Serper MCP Toolkit 是一个将 Serper 搜索 API 封装为 MCP 服务的工具包，提供网页、图片、视频、新闻、学术、购物、专利等搜索以及网页抓取能力，并支持 STDIO/HTTP/SSE 多种运行方式，适用于在 AI 应用与 MCP 客户端中集成实时搜索和结构化数据获取。

  • by Joey-Kot
  • Python
  • 本地部署
  • 更新于
在 GitHub 查看 Joey-Kot/serper-toolkit

使用说明

Serper MCP Toolkit

A rebuilt Serper MCP server based on the firecrawl-toolkit architecture.

Highlights

  • Unified async request layer (httpx + connection pooling + retries)
  • Compact JSON responses (single-line) for lower token usage
  • Stable mapped response schemas (no upstream raw passthrough)
  • Multi-page aggregation for non-image search endpoints (10 results/page granularity)
  • Image search search_num normalization to 10 or 100
  • Multi-transport startup (STDIO / HTTP / SSE, mutually exclusive)
  • Process lock to prevent duplicate server instances

Tools

  • serper-aggregated-search
  • serper-general-search
  • serper-image-search
  • serper-video-search
  • serper-place-search
  • serper-maps-search
  • serper-reviews-search
  • serper-news-search
  • serper-lens-search
  • serper-scholar-search
  • serper-shopping-search
  • serper-patents-search
  • serper-scrape

Core Parameter Rules

  • Search tools do not expose page; internal pagination is automatic.
  • Non-image endpoints:
    • search_num is clamped to 1..100
    • then rounded up to nearest multiple of 10 (25 -> 30)
  • Image endpoint:
    • 1..10 -> 10, 11..100 -> 100
  • Maps endpoint:
    • when aggregation requires page > 1 and query mode is used, ll is required.

Environment Variables

API and network

  • SERPER_API_KEY
  • SERPER_HTTP2 (0/1)
  • SERPER_MAX_CONNECTIONS
  • SERPER_KEEPALIVE
  • SERPER_MAX_CONCURRENT_REQUESTS
  • SERPER_MAX_WORKERS
  • SERPER_RETRY_COUNT
  • SERPER_RETRY_BASE_DELAY
  • SERPER_ENDPOINT_CONCURRENCY (JSON, e.g. {"search":10,"scrape":2})
  • SERPER_ENDPOINT_RETRYABLE (JSON, e.g. {"scrape":false})

Transport selection (exactly one must be enabled)

  • SERPER_MCP_ENABLE_STDIO
  • SERPER_MCP_ENABLE_HTTP
  • SERPER_MCP_ENABLE_SSE

Transport host/port

  • SERPER_MCP_HTTP_HOST
  • SERPER_MCP_HTTP_PORT
  • SERPER_MCP_SSE_HOST
  • SERPER_MCP_SSE_PORT
  • SERPER_MCP_HOST (fallback)
  • SERPER_MCP_PORT (fallback)

Lock file

  • SERPER_MCP_LOCK_FILE (default /tmp/serper_mcp.lock on Unix)

MCP Client Example

Use uvx to run the Python MCP server. uvx creates an isolated environment and installs package dependencies automatically:

{
  "mcpServers": {
    "serper-toolkit": {
      "command": "uvx",
      "args": ["serper-toolkit"],
      "env": {
        "SERPER_API_KEY": "<your-serper-api-key>",
        "SERPER_MCP_ENABLE_STDIO": "1"
      }
    }
  }
}

Go CLI

The Go CLI is located in the cli directory. It is a standalone command-line client named serper, separate from the Python MCP server.

The CLI reads the API key only from SERPER_KEY:

export SERPER_KEY="<your-serper-api-key>"

Build From Source

Requires Go 1.27.3 or later. CI reads the Go version from cli/go.mod.

Build for the current platform:

cd cli
go test ./...
go build -o serper .

Run it directly after building:

./serper --help

Build all release targets locally:

cd cli
mkdir -p dist

for target in windows/amd64 windows/arm64 linux/amd64 linux/arm64 darwin/amd64 darwin/arm64; do
  GOOS=${target%/*}
  GOARCH=${target#*/}
  binary=serper
  package="serper_${GOOS}_${GOARCH}.tar.gz"

  if [ "$GOOS" = windows ]; then
    binary=serper.exe
    package="serper_${GOOS}_${GOARCH}.zip"
  fi

  build_dir="build/${GOOS}_${GOARCH}"
  mkdir -p "$build_dir"
  CGO_ENABLED=0 GOOS=$GOOS GOARCH=$GOARCH go build -trimpath -ldflags="-s -w" -o "$build_dir/$binary" .

  if [ "$GOOS" = windows ]; then
    (cd "$build_dir" && zip -q "../../dist/$package" "$binary")
  else
    tar -C "$build_dir" -czf "dist/$package" "$binary"
  fi

  (cd dist && sha256sum "$package" > "$package.sha256")
done

(cd dist && find . -maxdepth 1 -type f ! -name '*.sha256' ! -name 'SHA256SUMS' -printf '%f\n' | sort | xargs sha256sum > SHA256SUMS)

CLI Commands

The CLI exposes the same 13 tool surfaces as subcommands:

serper aggregated --query "AI news" --search-num 20 --country US --language en --search-time day --timeout 120
serper general --query "AI news" --search-num 10 --country US --language en --search-time week
serper image --query "serper logo" --search-num 10
serper video --query "web scraping tutorial" --search-num 10
serper place --query "coffee" --location "Berlin,Germany"
serper maps --query "coffee" --ll "@52.5200,13.4050,14z" --search-num 20
serper reviews --cid "<google-cid>" --search-num 10 --sort-by newest
serper news --query "OpenAI" --search-time day
serper lens --image-url "https://example.com/image.jpg"
serper scholar --query "large language models"
serper shopping --query "mechanical keyboard"
serper patents --query "battery management"
serper scrape --output example --url "https://www.example.com"

Search-style commands output compact single-line JSON with the Python toolkit schema:

{"success":true,"meta":{"requested_search_num":10,"effective_search_num":10,"pages_fetched":1,"result_count":1,"credits":1},"data":{"organic":[]},"credits":1}

CLI Search Parameters

Shared search parameters:

  • --query (required for aggregated, general, image, video, place, maps, news, scholar, shopping, patents): Search keywords.
  • --search-num (optional): Number of results, range 1-100. Defaults to 20 for aggregated, and 10 for other search subcommands.
  • --country (optional): Country name or ISO code. Default is US where the Serper endpoint supports country targeting.
  • --language (optional): Language code, such as en.
  • --search-time (optional): One of hour, day, week, month, year. Supported by aggregated, general, image, video, and news.
  • --timeout (optional): Request timeout in seconds. Must be > 0. Default is 120.

Subcommand-specific parameters:

  • place: --location is optional.
  • maps: --ll, --place-id, and --cid are optional. When query-mode maps aggregation fetches more than one page, --ll is required.
  • reviews: at least one of --fid, --cid, or --place-id is required. --sort-by is optional.
  • lens: --image-url is required.

Normalization follows the Python toolkit:

  • Non-image endpoints clamp --search-num to 1-100, then round up to the nearest multiple of 10.
  • Image search uses 10 for requests 1-10, and 100 for requests 11-100.
  • Search tools do not expose a page parameter; pagination is automatic.

CLI Scrape Usage

Scrape a page and save the markdown export as example.md in the current directory:

serper scrape --output example --url "https://www.example.com" --timeout 120

Save the markdown export to a specific directory:

serper scrape --output example --path ./exports --url "https://www.example.com"

Scrape command parameters:

  • --output (required): Export name. The CLI writes <output>.md.
  • --path (optional): Directory where the markdown export is saved. Supports absolute and relative paths. Defaults to the current directory. If the directory does not exist, the CLI tries to create it before scraping.
  • --url (required): Target URL to scrape.
  • --include-markdown (optional): Request markdown content. Default is true.
  • --timeout (optional): Request timeout in seconds. Must be > 0. Default is 120.

Scrape output:

  • On success, stdout is true, and the CLI writes <output>.md.
  • On failure, stdout is false followed by the error reason, and no file is created or overwritten.

The generated markdown file uses this structure:

## title:
## description:
## url:
## credits:

---

markdown content

CI/CD

Two GitHub Actions workflows are included:

  • CI (.github/workflows/ci.yml)

    • Trigger: push/PR on main and dev, plus manual workflow_dispatch
    • Python matrix: 3.12, 3.13
    • Steps: alias validation, unit tests, build, twine check
    • Go CLI: tests, six-platform builds, and upload to the Latest GitHub Release on push/manual runs

  • Publish to PyPI (.github/workflows/publish-pypi.yml)

    • Trigger: push on main (automatic), plus manual workflow_dispatch
    • Preflight: alias validation, tests, build, twine check, version-not-on-PyPI check
    • Publish: trusted publishing via OIDC to PyPI

Required repository settings

  1. Create GitHub Environment: pypi
  2. Configure PyPI Trusted Publisher for this repo/workflow/branch:
    • Workflow file: .github/workflows/publish-pypi.yml
    • Branch: main
  3. Ensure Actions permissions allow OIDC (id-token: write in workflow is already set)

Development Validation

python3 scripts/validate_country_aliases.py
python3 -m unittest discover -s tests -p "test_*.py"
python3 -m build
python3 -m twine check dist/*

License

GNU General Public License v3.0 or later (GPL-3.0-or-later). See LICENSE for details. Third-party dependency notices and complete license texts are available in THIRD_PARTY_NOTICES.txt and THIRD_PARTY_LICENSES/.

相关 MCP