Content
<div align="center">
# Tool List
> *"You switched Claude Code to a Chinese LLM — and it can't browse the web anymore."*
[](LICENSE)
[](https://modelcontextprotocol.io)
[](#installation)
<br>
**One command to equip your agent with Doubao Search: ByteDance’s exclusive info source, long-form snippets up to 2000 chars, publish time down to the second, and 500 free searches/month.**
<sub>Doubao Search is a web information service built by Volcengine for AI agents, open to enterprises and developers since Jul 2026.</sub>
<br>
[Why you need it](#why-you-need-it) · [Effects](#effects) · [Installation](#installation) · [Tool parameters](#tool-parameters) · [Optional AI layer](#optional-ai-layer)
<br>
[English](#english)
</div>
---
## Why you need it
Claude Code's built-in WebSearch is an Anthropic **server-side** tool. Point `ANTHROPIC_BASE_URL` at any Anthropic-compatible endpoint (Doubao Seed, Kimi, GLM, DeepSeek...) and your agent silently goes offline.
## Effects
Here are real returns (raw output excerpts):
```
20 results, top 10 shown:
[1] "一人公司"与"手搓经济"，两个新词寓春意-新华网
Source: 新华网 | 2026-03-19T09:00:22+08:00
URL: http://www.xinhuanet.com/tech/20260319/259b0a6eb28c...
As AI technology advances rapidly, concerns about "job displacement" follow closely. Recently popular terms "一人公司" and "手搓经济" reveal another side of technological progress... (long-form snippet)
```
Some tested details:
- **Long-form snippets** (up to 2000 chars of article body per result) — consumable directly, no second crawl needed
- **Publish time down to the second** + source name on every result
- **Exclusive ByteDance sources**: Toutiao articles, Douyin Baike (top-ranked for entity lookups)
- **Cross-language**: English queries return first-party sources (official docs, changelogs)
- **`ContentTokenCount` per result** — budget your context window precisely
## Installation
### Step 1: Get an API Key
Go to [Volcengine Doubao Search Console](https://console.volcengine.com/search-infinity/web-search-exp) to enable the service and create an API Key. **500 free searches/month**, pay-as-you-go beyond.
### Step 2: Add to your Agent
**Claude Code** (one command):
```bash
claude mcp add huashu-doubao-search \
-e DOUBAO_SEARCH_API_KEY=your-key \
-- npx -y github:alchaincyf/huashu-doubao-search
```
Add `--scope user` for all projects.
**Cursor / Codex / other MCP clients** (config file):
```json
{
"mcpServers": {
"huashu-doubao-search": {
"command": "npx",
"args": ["-y", "github:alchaincyf/huashu-doubao-search"],
"env": {
"DOUBAO_SEARCH_API_KEY": "your-key"
}
}
}
}
```
**Easiest way**: share this repo with your agent and say:
```
Help me install this MCP: https://github.com/alchaincyf/huashu-doubao-search
```
### Install from source
```bash
git clone https://github.com/alchaincyf/huashu-doubao-search.git
cd huashu-doubao-search
npm install && npm run build
claude mcp add huashu-doubao-search \
-e DOUBAO_SEARCH_API_KEY=your-key \
-- node /absolute/path/huashu-doubao-search/dist/index.js
```
## Tool parameters
The agent now has a `doubao_search` tool:
| Parameter | Type | Default | Description |
|---|---|---|---|
| `query` | string | — | Search query, natural language works well |
| `count` | int 1-20 | 10 | Number of results |
| `snippet_length` | int 50-2000 | 600 | Max characters per snippet, adjust for in-depth reading |
| `images` | int 0-3 | 0 | Max images per result (CDN links, high bandwidth) |
| `max_age_days` | int 1-365 | off | Time filter: exclude results older than N days (no timestamp: keep). Pure code implementation, no extra cost |
## Optional AI layer
After retrieving search results, you can let a smaller model further process them. This layer is **optional, not default behavior**, with three switches in your control:
1. Without `ARK_API_KEY`, this layer doesn’t exist, behavior identical to plain search.
2. With `ARK_API_KEY`, enhancement capabilities registered but still output raw results.
3. Only when agent explicitly passes `max_tokens` or calls `doubao_cross_check` tool, AI processing occurs.
Processing layer only **filters and compresses**, not interpreting or concluding — source name, publish time, URL preserved, judgment left to your main model.
### Enable
Add two environment variables to installation command:
```bash
claude mcp add huashu-doubao-search \
-e DOUBAO_SEARCH_API_KEY=your-search-key \
-e ARK_API_KEY=your-ark-key \
-- npx -y github:alchaincyf/huashu-doubao-search
```
Defaults to `doubao-seed-2-0-lite-260215` (Seed 2.0 Lite, $0.6/million tokens, ~zero cost per process). Override with `ARK_MODEL` or `ARK_BASE_URL`.
### Capability 1: Token budgeter (`doubao_search` `max_tokens` parameter)
Search API includes `ContentTokenCount` per result. Use this parameter: pass `max_tokens: 800`, 10 long-form results compressed to ~800 tokens before entering your context — irrelevant results filtered out, preserving source, time, URL, and key sentences. If AI processing fails, automatically degrade to raw results.
### Capability 2: Multi-source cross-check (`doubao_cross_check` tool)
Given a question or statement to verify, it completes in one call: generates 3-4 search queries from different angles → parallel search, URL deduplication → structured report: **verification conclusion, source consensus, source discrepancies (each party’s statement + publish time), source list, freshness caveat**. Hard rules: only use searched content, no model knowledge introduced, transparent if sources insufficient.
This is a heavy tool (multi-search + two LLM calls, ~several dozen seconds), suitable for verifying rumors, breaking news with conflicting sources, or any fact worth multi-verification. For routine searches, use `doubao_search`.
## English
> *"You switched Claude Code to a Chinese LLM — and it can't browse the web anymore."*
Claude Code's built-in WebSearch is an Anthropic **server-side** tool. Point `ANTHROPIC_BASE_URL` at any Anthropic-compatible endpoint (Doubao Seed, Kimi, GLM, DeepSeek...) and your agent silently goes offline.
**huashu-doubao-search** plugs that gap with [Doubao Search](https://console.volcengine.com/search-infinity/web-search-exp) — the search API Volcengine built for AI agents, now open to developers with **500 free searches/month**:
- **Long-form snippets** (up to 2000 chars of article body per result) — consumable directly, no second crawl needed
- **Publish time down to the second** + source name on every result
- **Exclusive ByteDance sources**: Toutiao articles, Douyin Baike (top-ranked for entity lookups)
- **Cross-language**: English queries return first-party sources (official docs, changelogs)
- **`ContentTokenCount` per result** — budget your context window precisely
Install (Claude Code):
```bash
claude mcp add huashu-doubao-search \
-e DOUBAO_SEARCH_API_KEY=your-key \
-- npx -y github:alchaincyf/huashu-doubao-search
```
Or add the JSON config above to any MCP client. Tool: `doubao_search(query, count, snippet_length, images, max_age_days)`.
**Optional AI layer** (strictly opt-in — without `ARK_API_KEY` the server behaves exactly like the plain version): set `ARK_API_KEY` (Volcengine Ark) to unlock a `max_tokens` context-budget compressor on `doubao_search` (filter + compress only, sources/URLs/timestamps preserved, no conclusions added) and a `doubao_cross_check` tool (fans out 3-4 query angles, dedupes sources, returns a structured report: verdict / consensus / discrepancies / source list / freshness caveat). Powered by Doubao Seed 2.0 Lite by default (`ARK_MODEL` to override).
MIT License © [花叔 Huashu](https://github.com/alchaincyf)
<div align="center">
<sub>Author’s other projects · also by 花叔</sub>
[](https://github.com/alchaincyf/fanbox)
[女娲.skill — Distill anyone’s mindset](https://github.com/alchaincyf/nuwa-skill) · [达尔文.skill — Let Skills evolve infinitely](https://github.com/alchaincyf/darwin-skill) · [huashu-design — HTML native design skill](https://github.com/alchaincyf/huashu-design)
</div>
