io.github.rsdouglas/janee

编码与调试

by rsdouglas

面向 AI agents 的安全 secrets 代理,统一管理 API keys,避免代理直接接触原始凭证。

什么是 io.github.rsdouglas/janee

面向 AI agents 的安全 secrets 代理,统一管理 API keys,避免代理直接接触原始凭证。

README

Janee 🔐

Secrets management for AI agents via MCP

npm version npm downloads License: MIT GitHub stars

Your AI agents need API access to be useful. But they shouldn't have your raw API keys. Janee sits between your agents and your APIs — injecting credentials, enforcing policies, and logging everything.

✨ Features

🔒 Zero-knowledge agentsAgents call APIs without ever seeing keys
📋 Full audit trailEvery request logged with timestamp, method, path, status
🛡️ Request policiesAllow/deny rules per capability (e.g., read-only Stripe)
⏱️ Session TTLsTime-limited access with instant revocation
🔌 Works with any MCP clientClaude Desktop, Cursor, OpenClaw, and more
🏠 Local-firstKeys encrypted on your machine, never sent to a cloud
🖥️ Exec modeRun CLI tools with injected credentials — agents never see the keys
🤖 GitHub App authShort-lived tokens for autonomous agents — no static PATs
🐦 Twitter/X OAuth 1.0aPer-request OAuth signing — 4 secrets stay encrypted
☁️ AWS SigV4Sign AWS API requests server-side — SES, S3, and more
🔧 Automatic git authgit push/pull just works when credentials include GitHub tokens

The Problem

AI agents need API access to be useful. The current approach is to give them your keys and hope they behave.

  • 🔓 Agents have full access to Stripe, Gmail, databases
  • 📊 No audit trail of what was accessed or why
  • 🚫 No kill switch when things go wrong
  • 💉 One prompt injection away from disaster

The Solution

Janee is an MCP server that manages API secrets for AI agents:

  1. Store your API keys — encrypted locally in ~/.janee/
  2. Run janee serve — starts MCP server
  3. Agent requests access — via execute MCP tool
  4. Janee injects the real key — agent never sees it
  5. Everything is logged — full audit trail

Your keys stay on your machine. Agents never see them. You stay in control.


Configure Once, Use Everywhere

Set up your APIs in Janee once:

yaml
services:
  stripe:
    baseUrl: https://api.stripe.com
    auth: { type: bearer, key: sk_live_xxx }
  github:
    baseUrl: https://api.github.com
    auth: { type: bearer, key: ghp_xxx }
  openai:
    baseUrl: https://api.openai.com
    auth: { type: bearer, key: sk-xxx }

Now every agent that connects to Janee can use them:

  • Claude Desktop — access your APIs
  • Cursor — access your APIs
  • OpenClaw — access your APIs
  • Any MCP client — access your APIs

No more copying keys between tools. No more "which agent has which API configured?" Add a new agent? It already has access to everything. Revoke a key? Update it once in Janee.

One config. Every agent. Full audit trail.


Quick Start

Install

bash
npm install -g @true-and-useful/janee

Initialize

bash
janee init

This creates ~/.janee/config.yaml with example services.

Add Services

Option 1: Interactive (recommended for first-time users)

bash
janee add

Janee will guide you through adding a service:

code
Service name: stripe
Base URL: https://api.stripe.com
Auth type: bearer
API key: sk_live_xxx

✓ Added service "stripe"

Create a capability for this service? (Y/n): y
Capability name (default: stripe): 
TTL (e.g., 1h, 30m): 1h
Auto-approve? (Y/n): y

✓ Added capability "stripe"

Done! Run 'janee serve' to start.

Using an AI agent? See Non-interactive Setup for flags that skip prompts, or the agent-specific guides below.

Option 2: Edit config directly

Edit ~/.janee/config.yaml:

yaml
services:
  stripe:
    baseUrl: https://api.stripe.com
    auth:
      type: bearer
      key: sk_live_xxx

capabilities:
  stripe:
    service: stripe
    ttl: 1h
    autoApprove: true

Add CLI tools (exec mode)

Some tools need credentials as environment variables, not HTTP headers. Exec mode handles this:

bash
janee add twitter --exec \
  --key "tvly-xxx" \
  --allow-commands "bird,tweet-cli" \
  --env-map "TWITTER_API_KEY={{credential}}"

Now agents can run CLI tools through Janee without ever seeing the API key:

typescript
// Agent calls janee_exec tool
janee_exec({
  capability: "twitter",
  command: ["bird", "post", "Hello world!"],
  cwd: "/home/agent/project",  // optional working directory
  reason: "User asked to post a tweet"
})

Janee spawns the process with TWITTER_API_KEY injected, runs the command, and returns stdout/stderr. The credential never enters the agent's context.

Key flags:

  • --exec — configure as exec-mode (CLI wrapper instead of HTTP proxy)
  • --allow-commands — whitelist of allowed executables (security)
  • --env-map — map credentials to environment variables
  • --work-dir — working directory for the subprocess
  • --timeout — max execution time (default: 30s)

Git operations (automatic HTTPS auth)

When using exec mode with GitHub credentials, Janee automatically handles git authentication. No extra configuration needed — git push, git pull, and git clone just work:

yaml
capabilities:
  - name: git-ops
    service: github
    mode: exec
    allowCommands: [git]
    env:
      GH_TOKEN: "{{credential}}"
typescript
// Agent can push code without ever seeing the token
janee_exec({
  capability: "git-ops",
  command: ["git", "push", "origin", "main"],
  cwd: "/workspace/my-repo"
})

Janee detects git commands with GH_TOKEN/GITHUB_TOKEN in the environment and creates a temporary askpass script for HTTPS authentication. The script is cleaned up automatically after the command completes.

Add GitHub App auth (for autonomous agents)

Static tokens are risky for long-running agents. GitHub App auth generates short-lived installation tokens on demand — no long-lived PATs required.

Option 1: Use create-gh-app (recommended)

bash
npx @true-and-useful/create-gh-app create my-agent --owner @me
# Opens browser → creates app → saves credentials locally

# Install the app on your repos
# https://github.com/apps/my-agent/installations/new

# Register with Janee in one command
npx @true-and-useful/create-gh-app janee-add my-agent

Done. Your agent now gets short-lived GitHub tokens through Janee's MCP proxy.

Option 2: Manual setup

bash
janee add github-app \
  --auth-type github-app \
  --app-id 123456 \
  --pem-file /path/to/private-key.pem \
  --installation-id 789

Or via config:

yaml
services:
  github:
    baseUrl: https://api.github.com
    auth:
      type: github-app
      appId: "123456"
      pemFile: /path/to/private-key.pem
      installationId: "789"

How it works: When an agent requests access, Janee signs a JWT with the app's private key, exchanges it for a 1-hour installation token via GitHub's API, and caches the token until expiry. The agent never sees the private key — only the short-lived token reaches the API.

Start the MCP server

bash
janee serve

Use with your agent

Agents that support MCP (Claude Desktop, Cursor, OpenClaw) can now call the execute tool to make API requests through Janee:

typescript
// Agent calls the execute tool
execute({
  capability: "stripe",
  method: "GET",
  path: "/v1/balance",
  reason: "User asked for account balance"
})

Janee decrypts the key, makes the request, logs everything, and returns the response.


Integrations

Works with any agent that speaks MCP:


OpenClaw Integration

If you're using OpenClaw, install the plugin for native tool support:

bash
npm install -g @true-and-useful/janee
janee init
# Edit ~/.janee/config.yaml with your services

# Install the OpenClaw plugin
openclaw plugins install @true-and-useful/janee-openclaw

Enable in your agent config:

json5
{
  agents: {
    list: [{
      id: "main",
      tools: { allow: ["janee"] }
    }]
  }
}

Your agent now has these tools:

  • janee_list_services — Discover available APIs
  • janee_execute — Make API requests through Janee

The plugin spawns janee serve automatically. All requests are logged to ~/.janee/logs/.


MCP Tools

Janee exposes three MCP tools:

ToolDescription
list_servicesDiscover available APIs and their policies
executeMake an API request through Janee (HTTP proxy mode)
execRun a CLI command with injected credentials (exec mode)
manage_credentialView, grant, or revoke access to agent-scoped credentials
reload_configReload config from disk after adding/removing services (available when started with janee serve)

Agents discover what's available, then call APIs through Janee. Same audit trail, same protection.


Configuration

Config lives in ~/.janee/config.yaml:

yaml
server:
  host: localhost

services:
  stripe:
    baseUrl: https://api.stripe.com
    auth:
      type: bearer
      key: sk_live_xxx  # encrypted at rest

  github:
    baseUrl: https://api.github.com
    auth:
      type: bearer
      key: ghp_xxx

capabilities:
  stripe:
    service: stripe
    ttl: 1h
    autoApprove: true

  stripe_sensitive:
    service: stripe
    ttl: 5m
    requiresReason: true

Services = Real APIs with real keys
Capabilities = What agents can request, with policies

Supported auth types

TypeDescriptionExample
bearerBearer token in Authorization headerStripe, OpenAI, GitHub
basicHTTP Basic Auth (username + password)Internal APIs
hmac-bybitHMAC-SHA256 signing for BybitBybit exchange
hmac-okxHMAC-SHA256 + passphrase for OKXOKX exchange
hmac-mexcHMAC-SHA256 signing for MEXCMEXC exchange
headersCustom key-value headersNon-standard APIs
service-accountGoogle service account JSON keyGoogle Cloud
github-appShort-lived GitHub installation tokensGitHub API
oauth1a-twitterOAuth 1.0a per-request signingTwitter/X API v2
aws-sigv4AWS Signature V4 per-request signingSES, S3, and other AWS services

Twitter/X OAuth 1.0a

Janee computes OAuth 1.0a signatures (HMAC-SHA1) server-side, so your 4 Twitter secrets stay encrypted at rest and never enter the agent's context:

yaml
services:
  twitter:
    baseUrl: https://api.x.com
    auth:
      type: oauth1a-twitter
      consumerKey: xxx        # encrypted at rest
      consumerSecret: xxx     # encrypted at rest
      accessToken: xxx        # encrypted at rest
      accessTokenSecret: xxx  # encrypted at rest

capabilities:
  twitter:
    service: twitter
    ttl: 1h
    autoApprove: true

Or use the built-in template:

bash
janee add twitter

AWS SigV4

Janee computes AWS Signature V4 (HMAC-SHA256) per-request, keeping your access keys encrypted at rest. Non-secret fields (region, awsService) stay in plain config:

yaml
services:
  aws-ses:
    baseUrl: https://email.us-east-1.amazonaws.com
    auth:
      type: aws-sigv4
      accessKeyId: AKIA...     # encrypted at rest
      secretAccessKey: xxx     # encrypted at rest
      region: us-east-1
      awsService: ses

capabilities:
  aws-ses:
    service: aws-ses
    ttl: 1h
    autoApprove: true

Built-in templates for common AWS services:

bash
janee add aws-ses    # Amazon SES
janee add aws-s3     # Amazon S3

Access control

Control which agents can use which capabilities:

yaml
server:
  host: localhost
  defaultAccess: restricted   # capabilities require explicit allowlist

capabilities:
  stripe:
    service: stripe
    ttl: 1h
    allowedAgents: ["agent-a", "agent-b"]   # only these agents can use it

  github:
    service: github
    ttl: 1h
    # no allowedAgents + defaultAccess: restricted → no agent can use this
  • defaultAccess: restricted — capabilities without an allowedAgents list are hidden from all agents
  • defaultAccess: open (default) — capabilities without an allowedAgents list are available to all agents
  • allowedAgents — per-capability list of agent names (matched against clientInfo.name from the MCP initialize handshake)

Credentials created by agents at runtime default to agent-only access — only the creating agent can use them unless it explicitly grants access via the manage_credential tool.

Exec mode capabilities

yaml
services:
  twitter:
    auth:
      type: bearer
      key: tvly-xxx

capabilities:
  twitter:
    service: twitter
    mode: exec
    allowCommands: ["bird", "tweet-cli"]
    envMap:
      TWITTER_API_KEY: "{{credential}}"
    ttl: 1h
    autoApprove: true

Exec-mode capabilities use janee_exec instead of execute. The credential is injected as an environment variable — the agent sees only stdout/stderr.

Runner hardening defaults in exec mode:

  • isolated minimal environment (no full host env inheritance)
  • temporary HOME per command
  • timeout kills the process group

Runner/Authority mode (for containers)

When agents run inside Docker containers, janee_exec on a remote host cannot access the container filesystem. The Runner/Authority architecture solves this:

  • Authority runs on the host: holds credentials, enforces policy, proxies API requests
  • Runner runs inside each container: serves MCP to the agent, forwards non-exec calls to the Authority, runs janee_exec locally
bash
# Host: start Authority (MCP + exec authorization on one port)
janee serve -t http -p 3100 --host 0.0.0.0 --runner-key "$JANEE_RUNNER_KEY"

# Container: start Runner (agent talks to this)
janee serve -t http -p 3200 --host 127.0.0.1 \
  --authority http://host.docker.internal:3100 --runner-key "$JANEE_RUNNER_KEY"

The agent only needs JANEE_URL=http://localhost:3200.

You can also run the Authority as a standalone process:

bash
janee authority --runner-key "$JANEE_RUNNER_KEY" --host 127.0.0.1 --port 9120

See the Runner/Authority guide for the full architecture, exec authorization flow, Docker Compose example, and troubleshooting.


Request Policies

Control exactly what requests each capability can make using rules:

yaml
capabilities:
  stripe_readonly:
    service: stripe
    ttl: 1h
    rules:
      allow:
        - GET *
      deny:
        - POST *
        - PUT *
        - DELETE *

  stripe_billing:
    service: stripe
    ttl: 15m
    requiresReason: true
    rules:
      allow:
        - GET *
        - POST /v1/refunds/*
        - POST /v1/invoices/*
      deny:
        - POST /v1/charges/*  # Can't charge cards
        - DELETE *

How rules work:

  1. deny patterns are checked first — explicit deny always wins
  2. Then allow patterns are checked — must match to proceed
  3. No rules defined → allow all (backward compatible)
  4. Rules defined but no match → denied by default

Pattern format: METHOD PATH

  • GET * → any GET request
  • POST /v1/charges/* → POST to /v1/charges/ and subpaths
  • * /v1/customers → any method to /v1/customers
  • DELETE /v1/customers/* → DELETE any customer

This makes security real: Even if an agent lies about its "reason", it can only access the endpoints the policy allows. Enforcement happens server-side.


CLI Reference

bash
janee init                    # Set up ~/.janee/ with example config
janee add                     # Add a service (interactive)
janee add stripe -u https://api.stripe.com -k sk_xxx  # Add with args
janee remove <service>        # Remove a service
janee remove <service> --yes  # Remove without confirmation
janee list                    # List configured services
janee list --json             # Output as JSON (for integrations)
janee search [query]          # Search service directory
janee search stripe --json    # Search with JSON output
janee cap list                # List capabilities
janee cap list --json         # List capabilities as JSON
janee cap add <name> --service <service>  # Add capability
janee cap edit <name>         # Edit capability
janee cap remove <name>       # Remove capability
janee serve                   # Start MCP server (stdio, default)
janee serve --transport http --port 9100  # Start with HTTP transport (for containers)
janee serve --authority https://janee.example.com --runner-key $JANEE_RUNNER_KEY  # Runner mode
janee authority --runner-key $JANEE_RUNNER_KEY  # Start authority API
janee logs                    # View audit log
janee logs -f                 # Tail audit log
janee logs --json             # Output as JSON
janee sessions                # List active sessions
janee sessions --json         # Output as JSON
janee revoke <id>             # Kill a session

Non-interactive Setup (for AI agents)

AI agents can't respond to interactive prompts. Use --*-from-env flags to read credentials from environment variables — this keeps secrets out of the agent's context window:

bash
# Bearer auth (Stripe, OpenAI, etc.)
janee add stripe -u https://api.stripe.com --auth-type bearer --key-from-env STRIPE_KEY

# HMAC auth (Bybit)
janee add bybit --auth-type hmac-bybit --key-from-env BYBIT_KEY --secret-from-env BYBIT_SECRET

# HMAC auth with passphrase (OKX)
janee add okx --auth-type hmac-okx --key-from-env OKX_KEY --secret-from-env OKX_SECRET --passphrase-from-env OKX_PASS

# GitHub App auth (short-lived tokens)
janee add github --auth-type github-app --app-id-from-env GH_APP_ID --pem-from-env GH_PEM --installation-id-from-env GH_INSTALL_ID

# Twitter/X OAuth 1.0a (per-request signing)
janee add twitter --consumer-key $TWITTER_CONSUMER_KEY --consumer-secret $TWITTER_CONSUMER_SECRET \
  --access-token $TWITTER_ACCESS_TOKEN --access-token-secret $TWITTER_ACCESS_TOKEN_SECRET

# AWS SigV4 (SES, S3, etc.)
janee add aws-ses --access-key-id $AWS_ACCESS_KEY_ID --secret-access-key $AWS_SECRET_ACCESS_KEY \
  --region us-east-1 --aws-service ses

When all required credentials are provided via flags, Janee:

  • Never opens readline (no hanging on stdin)
  • Auto-creates a capability with sensible defaults (1h TTL, auto-approve)

You can also edit ~/.janee/config.yaml directly if you prefer.


How It Works

code
┌─────────────┐      ┌──────────┐      ┌─────────┐
│  AI Agent   │─────▶│  Janee   │─────▶│  Stripe │
│             │ MCP  │   MCP    │ HTTP │   API   │
└─────────────┘      └──────────┘      └─────────┘
      │                   │
   No key           Injects key
                    + logs request
  1. Agent calls execute MCP tool with capability, method, path
  2. Janee looks up service config, decrypts the real key
  3. Makes HTTP request to real API with key
  4. Logs: timestamp, service, method, path, status
  5. Returns response to agent

Agent never touches the real key.

📐 Deep dive: See Architecture & Security Model for detailed diagrams, threat model, and comparison with alternatives.


Security

  • Encryption: Keys stored with AES-256-GCM
  • Agent identity: Derived from clientInfo.name in the MCP initialize handshake — no custom headers needed
  • Agent isolation: Each agent gets its own session with isolated identity (HTTP transport creates a Server+Transport per session)
  • Access control: Per-capability allowedAgents whitelist + server-wide defaultAccess policy
  • Credential scoping: Agent-created credentials default to agent-only
  • Audit log: Every request logged to ~/.janee/logs/
  • Sessions: Time-limited, revocable
  • Kill switch: janee revoke or delete config

Docker

Run Janee as a container — no local Node.js required:

bash
# Build
docker build -t janee .

# Run in HTTP mode
docker run -d -p 3000:3000 \
  -v ~/.janee:/root/.janee:ro \
  janee --transport http --port 3000 --host 0.0.0.0

Or use Docker Compose:

bash
mkdir -p config && cp ~/.janee/config.yaml config/
docker compose up -d

For Claude Desktop with Docker, see Docker docs.


Contributing

We welcome contributions! Please read CONTRIBUTING.md before submitting a PR — it includes the required PR checklist (tests, changelog, version bump, etc.).


License

MIT — Built by True and Useful LLC


Stop giving AI agents your keys. Start controlling access. 🔐

常见问题

io.github.rsdouglas/janee 是什么?

面向 AI agents 的安全 secrets 代理,统一管理 API keys,避免代理直接接触原始凭证。

相关 Skills

前端设计

by anthropics

Universal
热门

面向组件、页面、海报和 Web 应用开发,按鲜明视觉方向生成可直接落地的前端代码与高质感 UI,适合做 landing page、Dashboard 或美化现有界面,避开千篇一律的 AI 审美。

想把页面做得既能上线又有设计感,就用前端设计:组件到整站都能产出,难得的是能避开千篇一律的 AI 味。

编码与调试
未扫描171.2k

网页应用测试

by anthropics

Universal
热门

用 Playwright 为本地 Web 应用编写自动化测试,支持启动开发服务器、校验前端交互、排查 UI 异常、抓取截图与浏览器日志,适合调试动态页面和回归验证。

借助 Playwright 一站式验证本地 Web 应用前端功能,调 UI 时还能同步查看日志和截图,定位问题更快。

编码与调试
未扫描171.2k

网页构建器

by anthropics

Universal
热门

面向复杂 claude.ai HTML artifact 开发,快速初始化 React + Tailwind CSS + shadcn/ui 项目并打包为单文件 HTML,适合需要状态管理、路由或多组件交互的页面。

在 claude.ai 里做复杂网页 Artifact 很省心,多组件、状态和路由都能顺手搭起来,React、Tailwind 与 shadcn/ui 组合效率高、成品也更精致。

编码与调试
未扫描171.2k

相关 MCP Server

GitHub

编辑精选

by GitHub

热门

GitHub 是 MCP 官方参考服务器,让 Claude 直接读写你的代码仓库和 Issues。

这个参考服务器解决了开发者想让 AI 安全访问 GitHub 数据的问题,适合需要自动化代码审查或 Issue 管理的团队。但注意它只是参考实现,生产环境得自己加固安全。

编码与调试
89.7k

by Context7

热门

Context7 是实时拉取最新文档和代码示例的智能助手,让你告别过时资料。

它能解决开发者查找文档时信息滞后的问题,特别适合快速上手新库或跟进更新。不过,依赖外部源可能导致偶尔的数据延迟,建议结合官方文档使用。

编码与调试
60.2k

by tldraw

热门

tldraw 是让 AI 助手直接在无限画布上绘图和协作的 MCP 服务器。

这解决了 AI 只能输出文本、无法视觉化协作的痛点——想象让 Claude 帮你画流程图或白板讨论。最适合需要快速原型设计或头脑风暴的开发者。不过,目前它只是个基础连接器,你得自己搭建画布应用才能发挥全部潜力。

编码与调试
49.9k

评论