Agentic Vault
MCP server that lets AI agents call APIs without ever seeing the credentials, using a local encrypted vault and per-secret allowlist policies for HTTP requests and subprocess environment variables.
README
Agentic Vault
MCP server that lets AI agents call APIs without ever seeing the credentials.
Secrets live in a local encrypted vault. The server substitutes them at call time into outbound HTTP requests or subprocess environment variables, under a per-secret allowlist policy. The model only ever sees secret names, never values.
Homepage: https://agenticvault.madhoob.dev
Why
Letting an AI agent make authenticated calls usually means giving it the raw API key. That breaks least-privilege, pollutes transcripts and logs, and means one jailbreak or prompt-injection is enough to exfiltrate the token.
Agentic Vault splits the two: the agent picks a secret by name and a destination; the vault checks the policy and injects the value. The plaintext never crosses the tool boundary.
Install
npm install -g secretproxy
secretproxy init # creates the global vault, stores master password in the OS keychain
secretproxy add OPENROUTER_API_KEY sk-or-...
secretproxy policy set OPENROUTER_API_KEY --host openrouter.ai
secretproxy run # start the MCP server over stdio
Then point any MCP client (Claude Code, Cursor, Cline, Codex, Zed) at secretproxy run.
Tools exposed over MCP
| Tool | Purpose |
|---|---|
list_secrets |
Enumerate available secret names (no values) |
http_request |
Make an HTTP call with a secret injected into headers, query, or body |
run_command |
Run a subprocess with secrets injected as env vars |
scan_env_requirement |
Detect what env vars a project expects and match them to stored secrets |
Features
- Zero-plaintext injection — values substituted inside the vault, never in the model context
- Per-secret policy — allow-lists for HTTP hosts, commands, env vars; deny by default; optional wildcards (strict mode rejects them)
- AES-256-GCM vault with argon2id key derivation
- Encrypted audit trail — every call logged with policy decision, surface, outcome
- Scoped vaults — global defaults + per-project overrides
- OS-native password storage — macOS Keychain, libsecret, Windows Credential Manager
- Interactive TUI (
secretproxy tui) and local-only web UI (secretproxy ui) - Rate limiting with token buckets
- Zero telemetry — no outbound calls, local-only by design
Development
npm install
npm test # 328 tests across 45 files
npm run build
npm run typecheck
Architecture primer: src/vault/ owns encryption, src/mcp/ owns the MCP tool surface, src/policy/ owns allowlist enforcement, src/audit/ owns the append-only log.
License
AGPL-3.0-or-later for open-source use — see LICENSE.
If your use case is incompatible with AGPL's network-copyleft clause (embedding in a proprietary product, offering as a managed service without source disclosure, etc.), a commercial license is available — see COMMERCIAL-LICENSE.md.
Contact: haaamcar@gmail.com
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。
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。
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。
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。
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。
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。
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。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。