VaultMCP
An encrypted secret vault and MCP gateway that securely stores API keys and injects them into upstream MCP servers, so AI IDEs never see raw credentials.
README
<div align="center">
VaultMCP
Keep API keys out of your AI IDE.
VaultMCP is an encrypted secret vault and MCP gateway. You store provider credentials once. Your AI IDE connects to one endpoint. VaultMCP decrypts secrets only when calling upstream MCP servers — your agents and config files never see the raw keys.
<table> <tr> <td align="center"> <img src="https://github.com/user-attachments/assets/f77ed65a-8f28-4f00-8cda-026eec6fa7cd" width="420" /> </td> <td align="center"> <img src="https://github.com/user-attachments/assets/2c537b03-70e0-493b-8ba4-867601078b5c" width="420" /> </td> </tr> </table>
Built by Axiler Labs · Product: vaultmcp.dev
<img src="docs/assets/clients/claude.svg" alt="Claude Desktop" height="28" /> <img src="docs/assets/clients/vscode.svg" alt="VS Code" height="28" /> <img src="docs/assets/clients/cursor.svg" alt="Cursor" height="28" /> <img src="docs/assets/clients/windsurf.svg" alt="Windsurf" height="28" /> <img src="docs/assets/clients/zed.svg" alt="Zed" height="28" />
<sub>Works with any MCP client — Claude Desktop, VS Code, Cursor, Windsurf, Zed, and more.</sub>
| Product | Self-host | For | License |
|---|---|---|---|
| vaultmcp.dev | https://YOUR_HOST |
AI IDEs + MCP secrets | AGPL-3.0 |
</div>
How it works
<div align="center">
One vault. One MCP URL. Every IDE talks to VaultMCP — not to your raw AWS or GitHub keys.
</div>
- Sign in with GitHub and create a workspace on vaultmcp.dev.
- Store secrets once (encrypted at rest — see Encryption). Share with teammates when needed.
- Point your IDE at
https://vaultmcp.dev/mcp(or your self-hosted/mcp). - Call tools like
aws__…/github__…. Credentials are injected server-side — never pasted intomcp.json.
flowchart LR
subgraph clients [MCP clients]
Claude[Claude]
VSCode[VS Code]
Cursor[Cursor]
Windsurf[Windsurf]
Zed[Zed]
end
subgraph vaultmcp [VaultMCP]
API[Vault · OAuth · MCP gateway]
end
subgraph upstreams [Upstreams]
MCP[AWS · GitHub · other MCP]
end
Claude --> API
VSCode --> API
Cursor --> API
Windsurf --> API
Zed --> API
API -->|inject secrets| MCP
Features
- Write-only vault — list APIs, MCP responses, and audit logs never return secret values
- Server-side injection — decryption only happens inside the API when calling an upstream
- Private and shared secrets — per-user or workspace-wide visibility
- GitHub OAuth or personal tokens — browser login for IDEs, or
vmcp_…tokens inAuthorization - Namespaced tools — register upstreams once; call them as
provider__tool_name - Single public port — Docker Compose exposes port 80; API and UI stay on the internal network
- Optional CLI — inject shared workspace secrets into local
npm run dev(and similar) without sharing.envfiles
Quick start (local)
Need: Node.js 22+, pnpm 9, Docker (Postgres + Redis), and a GitHub OAuth App.
git clone https://github.com/Axiler-Lab/vaultmcp.git
cd vaultmcp
cp .env.example .env
Fill in at least: VAULT_MASTER_KEY, GITHUB_CLIENT_ID, GITHUB_CLIENT_SECRET, PUBLIC_URL, WEB_ORIGIN, DATABASE_URL, REDIS_URL. Local defaults are in .env.example.
GitHub OAuth App
| Field | Local (pnpm) | Docker Compose (port 80) |
|---|---|---|
| Homepage URL | http://localhost:5173 |
http://YOUR_SERVER_IP |
| Callback URL | http://localhost:3001/auth/github/callback |
http://YOUR_SERVER_IP/auth/github/callback |
The callback must match PUBLIC_URL.
docker compose up -d postgres redis
pnpm install
pnpm --filter @vaultmcp/shared build
pnpm db:generate # first time / after schema changes
pnpm db:migrate
pnpm dev:api # :3001
pnpm dev:web # :5173
Open http://localhost:5173 → sign in → create a workspace → add secrets → register an upstream (AWS example). Host-run AWS uvx upstreams need uv on your PATH.
Docker Compose
One public entrypoint on port 80. API and web are not published directly.
cp .env.example .env
# Set PUBLIC_URL, WEB_ORIGIN, GITHUB_*, VAULT_MASTER_KEY
# VITE_API_URL= # empty = same-origin
# COOKIE_SECURE=false # true behind HTTPS
docker compose up --build
| Surface | URL |
|---|---|
| Web UI | http://localhost/ |
| Health | http://localhost/health |
| MCP | http://localhost/mcp |
| OAuth discovery | http://localhost/.well-known/oauth-protected-resource |
Migrations run when the API container starts. Production notes: docs/DEPLOY.md.
Connect Cursor (or any MCP client)
Same config shape for Cursor, Claude Desktop, VS Code, Windsurf, Zed, and other MCP clients. Use OAuth (browser login) or a personal access token from the dashboard Connect tab. Never put provider secrets in client config.
OAuth (hosted)
{
"mcpServers": {
"vaultmcp": {
"url": "https://vaultmcp.dev/mcp"
}
}
}
Personal token (vmcp_… from Connect)
{
"mcpServers": {
"vaultmcp": {
"url": "https://vaultmcp.dev/mcp",
"headers": {
"Authorization": "Bearer vmcp_…"
}
}
}
}
Self-host: replace with https://YOUR_HOST/mcp. Local API: http://localhost:3001/mcp.
Cursor: ~/.cursor/mcp.json or .cursor/mcp.json. Other clients use their own MCP config path.
After auth: list_workspaces → use_workspace → call namespaced tools (github__…, aws__…).
Local env CLI (optional)
For shared team secrets in local apps (not MCP agents), use @vaultmcp-axiler/cli:
npx @vaultmcp-axiler/cli@latest login --token vmcp_… --url https://YOUR_HOST
npx @vaultmcp-axiler/cli@latest run -w your-slug -- npm run dev
Create a PAT with the Runtime env (CLI) preset (env scope only). MCP tokens and env tokens are not interchangeable. Prefer run over env so secrets stay in the child process.
Full guide: docs/CLI.md.
Encryption
Secret values are encrypted at rest with AES-256-GCM using envelope encryption:
- Each workspace has its own data encryption key (DEK)
- DEKs are wrapped by
VAULT_MASTER_KEY(kept out of the database) - GCM AAD binds ciphertext to the
workspaceId, so blobs cannot be swapped across workspaces - Decryption happens only server-side — for MCP upstream injection, or opt-in CLI export
- Secret names and other metadata are not encrypted (so the UI can list and authorize without decrypting)
Treat VAULT_MASTER_KEY as a root credential. Rotate provider secrets in the UI without changing IDE config.
Also: list APIs and audit logs never return plaintext values; viewers cannot invoke secret-backed tools or export runtime env.
Details: docs/CRYPTO.md.
Repository layout
apps/api Gateway, OAuth, REST, upstream proxy
apps/web Control plane UI
packages/shared Crypto helpers + shared schemas
packages/cli Local env injection CLI
deploy/ Reverse proxy config
docs/ Deploy, CLI, and crypto guides
examples/ Upstream walkthroughs
Why AGPL
VaultMCP is often run as a network service. AGPL-3.0-only means if you modify it and offer that service over a network, you must share the corresponding source. That keeps hosted forks honest while still allowing self-hosting and contribution.
Contributing
VaultMCP is an open-source project from Axiler Labs. We’d love help from the community — issues and pull requests are very welcome, especially around security, reliability, docs, and MCP compatibility.
Ways to help:
- Open an issue for bugs or clear feature ideas
- Send a pull request (small, focused changes are easiest to review)
- Improve docs and examples for common upstreams
By contributing, you agree your work is licensed under AGPL-3.0-only, same as the rest of the repo. Never commit secrets; use .env.example for variable names only.
Learn more at vaultmcp.dev · Axiler Labs: axiler.com.
License
Copyright © 2026 Axiler Labs.
VaultMCP is licensed under the GNU Affero General Public License v3.0 only (AGPL-3.0-only).
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。