deepseek-web-search

Web search and page fetching for AI agents, on top of DeepSeek's Search and Fetch APIs. An agent skill, plus a Cloudflare Worker you host yourself.

Source on GitHub Is it worth building? DeepSeek docs

What you get

Skill	A `SKILL.md` that teaches an agent the search operators, result scores, error handling, rate limits and cost, with a zero-dependency CLI.
Worker	Access tokens, optional KV cache, batch extract, Tavily-shaped `/search` and `/extract`, and an MCP endpoint.
Licence	MIT. Bring your own DeepSeek key.

Deploy your own

git clone https://github.com/CreateSun/deepseek-web-search
cd deepseek-web-search/worker
npm install
npx wrangler login
npx wrangler secret put DEEPSEEK_API_KEY
npx wrangler secret put ACCESS_TOKENS
npx wrangler deploy

Callers send Authorization: Bearer <token> and your key stays on the server. Callers may instead send their own key in X-DeepSeek-Key; set ALLOW_BYOK = "false" to turn that off.

Try the demo instance

The demo Worker holds no DeepSeek key. It only forwards a key you send, and it does not store or log it. Because of that you are trusting me with the key, so use a throwaway one, or deploy your own.

curl https://deepseek-web-search.1819840437.workers.dev/v1/search \
  -H "X-DeepSeek-Key: $DEEPSEEK_API_KEY" -H "Content-Type: application/json" \
  -d '{"query":"DeepSeek V4 site:github.com","count":5}'

Endpoints

`POST /v1/search`	DeepSeek's native search body
`POST /v1/fetch`	`{"url"}` returns Markdown
`POST /search`	Tavily-shaped: domains, `time_range`, `lang`, optional cited `answer`
`POST /extract`	Up to 20 URLs, partial failures reported
`POST /mcp`	Tools `web_search` and `web_fetch`

Cost and honesty

DeepSeek charges 5 CNY per 1,000 successful Search or Fetch requests, with a 5 QPS limit per account. This project adds a thin layer on that and does not resell it. It is a utility, not a product; the feasibility notes explain why.