An OpenCode-native MCP server that gives your agents real-time web search via DeepSeek's server-side web_search_20250305 tool. One tool call handles search, page fetch, decryption, and answer synthesis — no third-party search API required.
Based on lyumeng/websearch-deepseek (MIT). See Attribution.
DeepSeek's Anthropic-compatible endpoint implements a built-in web_search_20250305 tool type. When your OpenCode agent calls the web_search tool, this MCP server forwards the query to DeepSeek, which performs the entire search pipeline server-side:
Agent calls web_search("latest Rust version")
│
▼
MCP Server ──POST──▶ api.deepseek.com/anthropic/v1/messages
tools: [{ type: "web_search_20250305" }]
│
▼ (all server-side)
1. Search the web
2. Fetch relevant pages
3. Decrypt page content
4. Synthesize a detailed answer
│
▼
MCP Server ◀──response── { text answer + source URLs }
│
▼
Agent receives Markdown answer with cited sources
No SerpAPI. No Tavily. No Brave Search key. DeepSeek does the searching itself.
Sign up at platform.deepseek.com and create an API key.
Add the server to your project's opencode.json (or ~/.config/opencode/opencode.json for global):
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"deepseek-websearch": {
"type": "local",
"command": ["npx", "@kyaulabs/deepseek-websearch"],
"enabled": true,
"environment": {
"DEEPSEEK_API_KEY": "{env:DEEPSEEK_API_KEY}"
}
}
}
}The {env:DEEPSEEK_API_KEY} syntax reads from your environment — set it in .envrc (direnv) or your shell profile:
export DEEPSEEK_API_KEY=sk-xxxxxxxxxxxxxxxxRestart OpenCode. Your agent now has a web_search tool available. Ask anything that needs current information:
- "What's new in React 19?"
- "Search for the latest Node.js LTS release schedule"
- "Find the current DeepSeek API pricing"
The agent will automatically invoke web_search when it needs real-time data beyond its training cutoff.
| Variable | Required | Default | Description |
|---|---|---|---|
DEEPSEEK_API_KEY |
Yes | — | DeepSeek API key |
WEBSEARCH_API_KEY |
No | — | Fallback key variable name |
WEBSEARCH_MODEL |
No | deepseek-v4-flash |
deepseek-v4-flash (fast) or deepseek-v4-pro (powerful) |
WEBSEARCH_THINKING |
No | enabled |
enabled or disabled |
WEBSEARCH_MAX_TOKENS |
No | 32768 |
Max response tokens |
WEBSEARCH_BASE_URL |
No | https://api.deepseek.com/anthropic |
API base URL (for proxies) |
Prefer a config file over env vars? Create ~/.deepseek-websearch.json:
{
"apiKey": "sk-xxxxxxxxxxxxxxxx",
"model": "deepseek-v4-pro",
"thinking": "disabled",
"maxTokens": 16384
}Resolution order (each layer overrides the previous): defaults → JSON file → environment variables.
| Model | Speed | Cost | Use When |
|---|---|---|---|
deepseek-v4-flash |
Fast | Low | Daily searches (default) |
deepseek-v4-pro |
Slower | Higher | Deep research, complex queries |
Each search consumes ~8,000–15,000 DeepSeek API tokens (search + thinking + answer generation). Check DeepSeek pricing for current rates.
- Zero runtime dependencies beyond the official
@modelcontextprotocol/sdk - Official MCP SDK — proper capability negotiation, error envelopes, no hand-rolled JSON-RPC
- TypeScript strict mode with 90%+ test coverage (Vitest)
- Structured errors — rate-limit detection (429), network errors, API errors, cancellation
- Configurable — env vars, JSON config file, or per-call programmatic overrides
- AbortSignal support — searches are cancellable
- Clean module separation — import
searchWeb()directly from your own code if you don't need the MCP layer
npm install # install dependencies
npm test # run unit test suite (56 tests)
npm run test:integration # run live API tests (requires DEEPSEEK_API_KEY)
npm run build # compile TypeScript → dist/
npm run check # type-check without emittingThe search logic is framework-agnostic. You can import it without the MCP server:
import { searchWeb } from "@kyaulabs/deepseek-websearch/search";
const result = await searchWeb("latest TypeScript features");
console.log(result.textAnswer); // AI-generated answer
console.log(result.results); // SearchResult[] with title, url, pageAgeThis project is based on lyumeng/websearch-deepseek by @lyumeng, originally released under the MIT License.
The original project established the approach of using DeepSeek's Anthropic-compatible endpoint with the web_search_20250305 tool type for server-side web search via MCP. This version is an independent engineering rewrite with the following improvements:
- Official
@modelcontextprotocol/sdkreplaces hand-rolled JSON-RPC - TypeScript strict mode with comprehensive Vitest test suite (90%+ coverage)
- Structured error handling with actionable codes (rate-limit detection, invalid config)
- Env vars + optional JSON config file with merge cascade
- Bug fix: system prompt placed as top-level
systemparameter (correct Anthropic Messages API format)