gingsisi/AgentsHive
DescriptiShared knowledge cache for AI agents — cache-first search saves tokens and avoids redundant web searches. Cross-agent deduplication with trust scoring. Human Bridge for blocked/paywalled content. MCP-native (FastMCP), ChromaDB-backed. 3 tools: agenthive_search, agenthive_contribute, agenthive_stats.
README
🐝 AgentsHive
<h3 align="center"> <b>Shared knowledge architecture for AI agents</b><br> <i>Breaking limits through human data acquisition.</i> </h3>
<p align="center"> <b>Let agents say High Five.</b><br> 群蜂聚智。一拍即合。 </p>
<p align="center"> <a href="#-why">Why</a> · <a href="#-features">Features</a> · <a href="#-quick-start">Quick Start</a> · <a href="#-api">API</a> · <a href="#-tag-system">Tag System</a> · <a href="#-deploy">Deploy</a> </p>
🤔 Why
AI agents can't reach everywhere. Paywalled sites, geo-blocked content, login-walled forums — bots hit walls every day. But humans don't.
AgentsHive is a human-powered knowledge mesh for AI agents. Real people browse the web — their captures flow into a shared cache. Agents query the mesh instead of burning tokens on blocked searches. One agent finds. Every agent knows.
| Without AgentsHive | With AgentsHive |
|---|---|
| 🔥 Burn tokens on every search | ⚡ Cache check → if hit, $0 |
| 🚫 Bots blocked everywhere | 👤 Humans go where bots can't |
| 🔁 Same queries, same cost | 🤝 One agent scrapes, all benefit |
| 🔒 Siloed per-user | 🌐 Open protocol, anyone joins |
✨ Features
| Feature | Description |
|---|---|
| 🧠 Shared Memory | Agents cache search results. One finds, every agent reuses. Zero duplicate tokens. |
| 👤 Human Bridge | Browser extension lets humans feed paywalled/geo-blocked content into the mesh. |
| 🌍 Region Auto-Detect | Server auto-detects geo-context from URL (TLD + subdirectory). .co.jp → JP, .gov.hk → HK. No manual tagging needed. |
| 🔒 Triple PII Defense | 3 layers: agent-side strip → user review queue → server re-scan. Emails, phones, IDs, credit cards all caught. |
| 🏷️ Smart Tag System | 22 canonical tags, 200+ aliases (EN/ZH/JP). Gaming, health, hobby, family — plus HK gov, finance, AI/ML. Cross-category minimum for quality. |
| 🔓 Open Protocol | BCP v0.1 — any agent joins. Any developer builds a node. The protocol is the moat. |
| ⚡ JSON Config | Edit references/canonical_tags.json → git push → live. No redeploy, no downtime. mtime-based auto-reload. |
🏗️ How It Works
Agent → Mesh → Every Agent
🤖 Agent A searches "Elden Ring boss guide" → cache miss → scrapes web → contributes
🤖 Agent B searches same query → cache HIT → instant answer, 0 tokens
👤 Human sees geo-blocked Tabelog page → Human Bridge capture → mesh now has JP content
Contribution Schema (20 fields, 5 groups):
┌─ Core ─────────────┬──────────────────────┐
│ id │ wr_a1b2c3d4e5f6 │
│ query │ "Elden Ring攻略" │
│ content │ "Boss弱點係..." │
│ source_url │ https://... │
│ content_hash │ SHA-256 of content │
├─ Context ───────────┼──────────────────────┤
│ language │ zh (auto-detect) │
│ region │ JP (auto-detect) │
│ token_size │ 1250 (estimated) │
├─ Classification ────┼──────────────────────┤
│ tags │ gaming,hobby │
│ privacy_class │ public │
│ filtration_status │ scanned/redacted │
│ verification │ unverified→verified │
├─ Attribution ───────┼──────────────────────┤
│ contributor_id │ usr_998877 │
│ is_human_bridged │ true/false │
├─ Lifecycle ─────────┼──────────────────────┤
│ created │ Unix timestamp │
│ expires │ +30 days │
│ reproductions │ reuse count │
└─────────────────────┴──────────────────────┘
🚀 Quick Start
# Clone
git clone https://github.com/gingsisi/AgentsHive.git
cd AgentsHive
# Install
pip install -r requirements.txt
# Run (default port 15000)
python3 server.py
# → http://localhost:15000
# Get an API key (2-step email verification)
# Open http://localhost:15000 in your browser → sign up
# Search the mesh
curl "http://localhost:15000/search?q=disability+allowance&n=3" \
-H "X-API-Key: bc_YOUR_KEY"
# Contribute (requires 2+ cross-category canonical tags)
curl -X POST http://localhost:15000/contribute \
-H "X-API-Key: bc_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "HK immigration FAQ",
"content": "Visit visa extension requires...",
"source_url": "https://www.immd.gov.hk/eng/faq.html",
"tags": ["immigration", "hong-kong"]
}'
# Human Bridge capture (auto-sets is_human_bridged=true)
curl -X POST http://localhost:15000/bcp/v1/bridge-capture \
-H "X-API-Key: bc_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"query":"...","content":"...","source_url":"...","tags":["travel","food"]}'
🏷️ Tag System
22 canonical tags across 6 categories. Every contribution needs 2+ cross-category tags. Aliases handle the mapping — use any alias, server normalizes.
| Category | Tags | Example Aliases |
|---|---|---|
| Domain | welfare, tax, education, medical, law, finance, policy, housing, immigration | swd, cssa, ird, edb, asd, adhd, 港股 |
| Technical | tech-dev, tech-creative, ai-ml | godot, blender, zbrush, llm, python |
| Geo | hong-kong, china, international | hk, 香港, 大陸 |
| Lifestyle | travel, food, shopping, transport, lifestyle | japan, jr, fukuoka, 拉麵, 購物 |
| Consumer | gaming, health, hobby, family | rpg, steam, fitness, anime, 育兒, sen |
| Meta | temporary, evergreen | covid, basics |
Region auto-detect: If no region field provided, server detects from URL:
- TLD:
.co.jp→JP,.gov.hk→HK,.co.uk→UK - Subdirectory:
/ja-jp/→JP,/en-us/→US - Fallback:
Global
Tag config: Edit references/canonical_tags.json — structured by category with aliases + decay_days. JSON is mtime-cached for zero-overhead reload.
🔌 API
Public Endpoints
| Endpoint | Description |
|---|---|
GET / |
Landing page (5 languages: EN/中文/日本語/한국어) |
GET /health |
Health check |
GET /stats |
Cache statistics |
POST /api/keys/request-verification |
Step 1: email verification |
POST /api/keys/generate |
Step 2: verify code → get API key |
GET /tos |
Terms of Service |
GET /privacy |
Privacy Policy |
Authenticated (X-API-Key header)
| Endpoint | Description |
|---|---|
GET /search?q=...&n=3 |
Semantic search the mesh |
POST /contribute |
Contribute a web result |
POST /bcp/v1/bridge-capture |
Human Bridge capture (Layer 3 PII scan) |
GET /api/keys/stats |
Your key's usage stats |
Rate Limits (Beta)
| Tier | Search | Contribute | Price |
|---|---|---|---|
| Free | 10,000/min | 10,000/min | $0/mo (beta) |
| Pro | Unlimited | 50/min | Coming soon |
🌍 Multi-Language
The landing page supports 5 languages with instant switching, no page reload:
- 🇬🇧 English
- 🇭🇰 繁體中文
- 🇯🇵 日本語
- 🇰🇷 한국어
- 🇨🇳 简体中文
☁️ Deploy
# Railway (free tier)
# Just point to server.py — single FastAPI app
# Procfile: web: uvicorn server:app --host 0.0.0.0 --port $PORT
# Environment variables:
# BC_SYSTEM_KEY — system key for production
# RESEND_API_KEY — email service (Resend)
# PORT — server port (default 15000)
# Docker
docker build -t agenthive .
docker run -p 15000:15000 agenthive
🛡️ Privacy & Security
- All contributed content is user-contributed and unverified
- PII auto-stripped at 3 layers: agent → review → server re-scan
- API keys SHA-256 hashed in SQLite
- No personal data stored — queries cached by content hash, not user identity
- 13 PII patterns caught: emails, phones (HK/CN/US/UK/JP/TW/SG/KR), national IDs (HKID, SSN, NIN, NRIC, RRN, CPF), IPs, API keys, credit cards, passports
📦 Tech Stack
- FastAPI — REST API
- ChromaDB — Vector storage
- SQLite — API key management
- Sentence Transformers — Embeddings (English)
- Ollama — Chinese embeddings (nomic-embed-text)
🤝 Contributing
AgentsHive is open source under MIT License. See CONTRIBUTING.md.
<p align="center"> <b>🐝 Built for agents. Contributed by humans.</b><br> <i>Let agents say High Five. 群蜂聚智。一拍即合。</i> </p>
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。