
Keyblind
io.github.aarifmms由 GitHub 账号 aarifmms 发布v0.8.8更新于 Oct 24, 2026
Encrypted secrets vault for AI agents. 27 MCP tools resolve secrets at runtime, never leaking them.
概览
本地加密密钥保险库,通过 MCP 让 AI 助手存取、列出和解析密钥,而明文不会进入对话记录。
- 可以做什么
- 在本地加密保险库中存储、列出、解析和删除密钥,值不进入对话记录
- 对 .env 文件做沙箱与还原,让代理只看到确定性的假值
- 查看审计日志、生成 TOTP 验证码,创建或接收加密的限时分享链接
- 管理团队保险库、注册代理、记录签名操作并导出合规报告
- 查询死亡开关状态与 SSO/OIDC 认证状态
- 适用场景
- 适合让编码代理在需要 API 密钥、数据库连接串或令牌的项目中工作,并希望这些值在运行时解析,而不是被粘贴进对话或提交到 .env 文件。也适合需要审计记录、按密钥策略或代理访问审批的团队。
- 运行要求
- 以 npm 包 keyblind 形式通过 stdio 在本地运行,通常用 npx 或 keyblind start 启动,需先初始化保险库。未声明任何账号、API 密钥或环境变量。可选的 1Password、Bitwarden、AWS、GCP、Azure 后端需要各自对应的 CLI。团队保险库、审计日志、分享、TOTP、生物识别门禁和 SSO 等功能需要付费的 Pro 或 Team 许可证。
安装
在 SourceWeft 中
- 打开 控制台中的 Keyblind,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
Keyblind — Blind AI to Your Keys
Encrypted secrets vault with MCP for AI agents. One command configures every editor. Secrets resolved at runtime, never leaked to LLM conversations.
Why
Developers regularly leak API keys, passwords, and tokens to AI coding tools. 100,000+ LLM conversations with exposed secrets were found indexed by search engines in 2025.
AI agents read your .env files. They copy-paste secrets into conversations. They commit them accidentally. Keyblind stops this by keeping secrets encrypted at rest and resolving them at runtime — the plaintext value never touches the LLM transcript.
How It Works
Quick Start
That's it. After
keyblind quickstart, restart Claude Code. Then just say "list my keyblind secrets" or "use my OPENAI_API_KEY" — the AI agent resolves secrets at runtime without ever seeing them in the transcript.
MCP Server
Keyblind is MCP-first — it works with every AI tool that speaks the Model Context Protocol (Claude Code, Cursor, Copilot, Windsurf, Cline, Zed).
Setup (automatic)
This auto-configures Claude Code, Cline, and Zed to use Keyblind. Works from any directory.
For project-local editors (Cursor, Windsurf, Copilot, Cline project mode), run from your project root:
See editor-specific configs for details and manual setup.
Setup (manual)
Add a .mcp.json to your project root, or use claude mcp add:
With biometric gate (Touch ID required before secrets are resolved):
Session expires after 15 minutes. Requires Pro or Team license.
MCP Tools
Web Dashboard
Manage your secrets from a browser — no sign-in or pairing required. Open the local dashboard:
This starts the local server and opens http://localhost:4318/ in your browser. The dashboard is served same-origin with the REST API, so it works in every browser with no CORS, no pairing token, and no login. Features:
- View, add, copy, and delete secrets
- Audit log with full access history
- License management (tier, email, expiry)
- Pro/Team tier status
- Per-secret access policies and rotation reminders
- Agent action approvals (approve/deny queued agent writes)
- Central control plane (org-wide visibility + signed revocation)
A hosted control plane (hosted/) adds orgs, roles, groups, a signed
revocation + group-scope manifest, multi-provider SSO login (with org
auto-provisioning, multi-org switching, and IdP end-session revocation), and
short-lived agent credentials. See docs/enterprise.md (Phase 2).
Browser Extension
The Keyblind Chrome Extension detects and blocks secrets from being pasted into AI chat interfaces (Claude.ai, ChatGPT, Copilot).
Features:
- Detects 12+ API key formats (OpenAI, GitHub, Stripe, AWS, etc.)
- Intercepts paste events on AI chat sites
- Warning banner when secrets are detected
- Popup with vault connection status
Located in browser-extension/. Load as unpacked extension from chrome://extensions.
Pricing
Backends
Keyblind supports multiple secret backends:
Keyblind vs Cloak
Security
- AES-256-GCM encryption with PBKDF2 key derivation (600K iterations)
- Machine-identity-bound key — encryption key XOR-wrapped with machine fingerprint
- Zero network, zero telemetry — no cloud, no accounts, no analytics
- Vault stored at
~/.keyblind/with0700permissions - Deterministic sandbox fakes using HMAC-SHA256 per project + key name
Verifying Your Install
Releases are built and published from GitHub Actions with npm provenance, which attests the tarball was produced from this repository's CI rather than a developer machine.
To verify build provenance for a release, open the release page at https://github.com/aarifmms/keyblind/releases and check the "Provenance" and "Attestations" sections, which link the tarball to the exact commit and workflow run that built it.
- Small dependency graph — the CLI depends on three packages only (MCP SDK,
better-sqlite3, zod). Server-only deps (Stripe, Resend, the Vercel CLI) live in
the
hosted/anddashboard/packages and are not installed bynpm i -g keyblind. - No source maps — the published tarball ships compiled JavaScript with no embedded original source.
CLI Reference
Development
License
MIT
来源:README.md,提交 6622cbb
工具
0版本历史
2- v0.8.8最新Oct 21, 2026
- v0.1.4Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分