API Key Manager for AI Agents — 专为 AI Agent 设计的 API 密钥管理服务
English | 中文
Keyring is a lightweight, self-hosted service for managing API keys across platforms — built specifically for AI agents.
Instead of scattering API keys across .env files, config files, and environment variables, Keyring gives your agents a single, structured, searchable store for all their credentials.
| HashiCorp Vault | Infisical | Keyring | |
|---|---|---|---|
| Deploy | Complex (Go binary, storage backend) | Needs PostgreSQL | Single JSON file |
| Target | Enterprise teams | Dev teams | Solo developers / AI agents |
| MCP support | Via plugin | Via plugin | Native, with auto-discovery |
| Weight | Heavy | Medium | Light (JSON file, no DB) |
Keyring fills the gap between .env files (too primitive) and enterprise secret managers (too heavy).
- 8 MCP tools — agents can list, search, get, add, update, revoke keys, view version history, get summaries
- MCP auto-discovery —
/.well-known/mcp,/llms.txt, HTML<link>tags, HTTPLinkheader - Version control — every edit saves a snapshot; full history per key
- Soft delete — revoked keys stay in the store, can be reactivated
- Semi-structured data — models, tags, notes are flexible fields
- WebUI — visual management at port 5179
- Zero dependencies — JSON file storage, atomic writes, no database needed
- Two transport modes — HTTP (Streamable HTTP) for remote, stdio for local
git clone https://github.com/Leeson-Wong/Keyring.git
cd Keyring
cp config.example.json config.json
docker compose up -dVisit http://localhost:5179 for the WebUI.
MCP endpoint: http://localhost:5179/mcp
git clone https://github.com/Leeson-Wong/Keyring.git
cd Keyring
npm install
npm run build
cp config.example.json config.json
npm startFor Claude Desktop, Cursor, or any MCP client that supports stdio:
{
"mcpServers": {
"keyring": {
"command": "npx",
"args": ["tsx", "server/stdio.ts"],
"cwd": "/path/to/Keyring"
}
}
}{
"mcpServers": {
"keyring": {
"url": "http://YOUR_HOST:5179/mcp"
}
}
}Any MCP-compatible agent visiting the root URL will find the MCP endpoint via:
/.well-known/mcp— machine-readable JSON/llms.txt— LLM-readable markdown- HTML
<link>tags in the homepage - HTTP
Linkresponse header
| Tool | Description |
|---|---|
list_keys |
List keys, optionally filtered by platform / function / status |
get_key |
Get full details of a key (including the actual API key value) |
search_keys |
Full-text search across all fields |
add_key |
Add a new API key |
update_key |
Update a key (auto-saves version) |
revoke_key |
Soft-delete a key |
get_versions |
View version history |
get_summary |
Overview: counts, platforms, functions, tags |
| Method | Path | Description |
|---|---|---|
| GET | /api/keys |
List all keys |
| GET | /api/keys/:id |
Get a key |
| POST | /api/keys |
Create a key |
| PUT | /api/keys/:id |
Update a key |
| DELETE | /api/keys/:id |
Revoke a key (soft delete) |
| GET | /api/keys/:id/versions |
Version history |
| GET | /api/meta |
Platforms, functions, tags |
| POST | /mcp |
MCP endpoint |
| GET | /health |
Health check |
{
"id": "uuid",
"platform": "openai",
"function": "llm",
"endpoint": "https://api.openai.com/v1",
"models": ["gpt-4o", "gpt-4o-mini"],
"apiKey": "sk-...",
"name": "My OpenAI key",
"tags": ["production", "coding"],
"notes": "Main key for code generation",
"status": "active",
"createdAt": "2025-01-01T00:00:00Z",
"updatedAt": "2025-01-01T00:00:00Z",
"versions": []
}If you deploy Keyring on a local network (e.g. VMware internal IP, home LAN), there's no attack surface — the service is not exposed to the public internet. In this case, no authentication is needed. This is the default and recommended usage.
If you need to expose Keyring to the public internet, you must add authentication. Options:
- Reverse proxy with Basic Auth — simplest approach (Nginx / Caddy)
- API key header — check
Authorizationheader in a middleware - Reference implementation — see resume-maker for a working example with invite-code authentication, rate limiting, and MCP discovery
- Use a VPN / tunnel — WireGuard, Tailscale, or Cloudflare Tunnel to avoid direct exposure
⚠️ Never deploy Keyring to the public internet without authentication. Your API keys are stored in plaintext.
{
"port": 5179,
"dataPath": "./data"
}Or via environment variables:
PORT=5179 DATA_PATH=./data npm startKeyring/
├── server/ # Backend (Node.js + MCP SDK)
│ ├── index.ts # HTTP server + API routes + MCP discovery
│ ├── stdio.ts # stdio MCP transport
│ ├── storage.ts # JSON file storage engine
│ └── config.ts # Config loader
├── mcp/ # MCP server + tools
│ ├── create-server.ts # McpServer factory
│ └── tools/index.ts # 8 MCP tool definitions
├── src/ # Frontend (React 19 + Vite 7 + Tailwind CSS 4)
│ ├── types/keyring.ts # Shared types
│ └── ... # React components
├── data/ # Runtime data (gitignored)
│ └── keyring.json # Key store
├── Dockerfile
├── docker-compose.yml
└── package.json
- Backend: Node.js, TypeScript, MCP SDK, Zod
- Frontend: React 19, Vite 7, Tailwind CSS 4
- Storage: JSON file (atomic writes, no database)
- Deploy: Docker, docker-compose, or bare Node.js
Your Hermes / Claude / Cursor agent needs OpenAI, Anthropic, ElevenLabs keys. Instead of scattering them in config files, store them in Keyring and let agents query via MCP.
Multiple keys per platform? Keyring tracks them all with version history. When you rotate a key, the old version is preserved.
Deploy on a shared server. Everyone on the LAN can access the same keys via MCP or WebUI.
MIT
Keyring 是一个专为 AI Agent 设计的 API 密钥管理服务，轻量自托管，支持 MCP 协议，Agent 可以直接通过工具调用来查询、添加、更新密钥。
每个用 AI Agent 的人都有一堆 API Key：OpenAI、Anthropic、ElevenLabs、各种搜索 API……这些 key 散落在 .env、配置文件、环境变量里，Agent 找不到、人也记不清。
现有的方案：
- HashiCorp Vault — 企业级，太重
- Infisical — 需要数据库，部署复杂
- .env 文件 — 太原始，没有结构化、没有版本控制
Keyring 填的就是这个中间地带：比 .env 结构化，比 Vault 轻量，原生支持 MCP 让 Agent 直接用。
- 内网部署（推荐）：没有攻击面，不需要设置认证
- 公网部署：务必加认证。最简单的是 Nginx 反代 + Basic Auth，或者参考 resume-maker 的邀请码认证方案