Deepseek Search

DeepSeek Search

DeepSeek Search is a Claude Code plugin that wraps DeepSeek's Anthropic-compatible Web Search capability as a local Python CLI and exposes it through a Claude Code skill.

What it provides

• A deepseek-search CLI (Python + uv), published on PyPI.
• DeepSeek API configuration under ~/.deepseek.
• A Claude Code skill at plugins/deepseek_search/skills/deepseek_search/SKILL.md that tells Claude Code when and how to call the CLI.

Project layout

.claude-plugin/marketplace.json
.github/workflows/          # CI + publish-to-PyPI on `v*` tags
plugins/deepseek_search/    # Claude Code plugin manifest + skill
  .claude-plugin/plugin.json
  skills/deepseek_search/SKILL.md
src/deepseek_search.py      # CLI implementation
tests/test_deepseek_search.py
pyproject.toml
uv.lock

The Python package lives at the repository root; plugins/deepseek_search/ holds only the Claude Code plugin manifest and skill instructions.

Install

The recommended way to use the CLI is as a global uv tool:

uv tool install deepseek-search

Then deepseek-search is available on PATH everywhere.

From source (development)

git clone https://github.com/8DE4732A/deepseek-search.git
cd deepseek-search
uv sync --dev            # install dependencies
uv run pytest            # run tests (no API key required)

Configure DeepSeek

All configuration lives under ~/.deepseek and is managed through the config subcommand. There are no environment variables — settings come from the config files or one-off CLI flags.

deepseek-search config init                                    # write a default config.json
deepseek-search config set api_key YOUR_DEEPSEEK_API_KEY       # stored in ~/.deepseek/credentials.json (chmod 600)
deepseek-search config set model deepseek-v4-pro[1m]           # set any field
deepseek-search config get model                               # read a single value
deepseek-search config list                                    # show full effective config (api_key masked)

Configurable keys (type-coerced automatically):

Key	Stored in	Type
api_key	credentials.json	string (chmod 600)
base_url, model, web_search_tool_type, output_format	config.json	string
max_tokens, max_retries	config.json	int
timeout_seconds	config.json	float
max_uses, allowed_domains, blocked_domains	config.json (search.*)	int / comma-list

Advanced search flags (--model, --max-tokens, --timeout, --max-uses, --allowed-domain, --blocked-domain, …) override config.json for a single run without writing it back.

CLI usage

deepseek-search --json --query "DeepSeek Claude Code web search documentation"
deepseek-search "DeepSeek API Claude Code integration"        # bare positional query also works
deepseek-search search --markdown "DeepSeek API Claude Code integration"
deepseek-search --help

The JSON output includes:

• answer
• citations
• warnings
• usage
• model
• request_id
• stop_reason

Releasing / publishing

Releases are produced automatically by the Publish to PyPI GitHub Action whenever a v* git tag is pushed:

git tag v0.1.0
git push origin v0.1.0

Prerequisites (one-time, in the GitHub repo + PyPI):

1. Register the project as a Trusted Publisher on PyPI:
   • PyPI publishing environment: pypi
   • GitHub repository: 8DE4732A/deepseek-search
   • Workflow filename: publish.yml
   • Environment name: pypi
2. No PyPI API token is required — publishing uses OIDC.

The package version is derived from the git tag via setuptools-scm, so the workflow fails fast if the tag and the built version disagree.

License

MIT License. See LICENSE.