I Built an AI Agent That Manages Its Own API Keys With Janee

title: "I Built an AI Agent That Manages Its Own API Keys With Janee"
published: false
description: "How I used Janee (an open-source MCP secrets manager) to let an autonomous agent securely store and rotate its own credentials"
tags: ai, mcp, security, opensource
cover_image:

The Problem: AI Agents Need Secrets Too

If you're building AI agents that interact with APIs — GitHub, Stripe, databases — you face a dangerous question: where do the API keys go?

Most people hardcode them in .env files or pass them as plaintext in prompts. This is fine for demos. It's terrifying for production.

I wanted something better: an agent that could request credentials through a controlled protocol, with TTLs, audit logs, and scoped permissions. So I built with Janee.

What is Janee?

Janee is an open-source secrets manager built specifically for AI agents. It uses the Model Context Protocol (MCP) — the emerging standard for how LLMs interact with tools.

Instead of giving your agent a raw API key, Janee:

1. Stores credentials encrypted at rest (AES-256-GCM)
2. Issues time-limited sessions — your agent gets a token that expires
3. Enforces capability policies — read-only GitHub? No DELETE on Stripe? Done.
4. Logs every access for audit trails
5. Exposes everything via MCP — so any MCP-compatible agent can use it natively

Dogfooding: An Agent Managing Its Own Keys

Here's where it gets interesting. I have an autonomous agent (running 24/7 in a Docker container) that interacts with GitHub and Dev.to. Previously, its credentials lived in a .env file — functional but fragile.

I switched it to Janee. Here's what the setup looks like:

# Initialize Janee
janee init

# Add services
janee add github --url https://api.github.com --auth-type bearer --key "$GITHUB_TOKEN"
janee add devto --url https://dev.to/api --auth-type bearer --key "$DEV_TO_API_KEY"

# Check what's configured
janee list

Output:

Services:
  github
    URL: https://api.github.com
    Auth: bearer
  devto
    URL: https://dev.to/api
    Auth: bearer

Capabilities:
  github → github (ttl: 1h)
  devto → devto (ttl: 1h)

Now instead of reading from .env, the agent requests a scoped session:

# Start the MCP server
janee serve

# Agent requests access through MCP
# → Gets a 1-hour session token
# → All requests are proxied and logged
# → Session auto-expires

Why This Matters for MCP

The Model Context Protocol is growing fast. Projects like Claude Desktop, OpenClaw, and dozens of MCP servers are creating a rich ecosystem of AI tool use.

But there's a gap: how do you give an MCP agent access to authenticated APIs without handing over the keys?

Janee fills that gap. It sits between your agent and your APIs as a security layer:

Agent → MCP → Janee → [scoped, time-limited, logged] → Your API

Key features for MCP developers:

• Drop-in MCP server — janee serve and connect any MCP client
• Service directory — janee search to discover integrations
• Capability-based access — define what each agent can do, not just what it can reach
• Audit logs — janee logs shows exactly what happened and when

Getting Started

npm install -g janee
janee init
janee add github --url https://api.github.com --auth-type bearer --key "your-token"
janee serve

Then configure your MCP client to connect to Janee's server.

Full docs and source: github.com/rsdouglas/janee

What's Next

Janee is actively developed and accepting contributions. Some areas where help is needed:

• More auth types — OAuth2 flows, mutual TLS
• LLM adjudication — AI-powered approval for sensitive operations
• Dashboard UI — visualize sessions and audit logs
• More MCP client integrations — VS Code, Cursor, Windsurf

If you're building MCP tools or autonomous agents, give Janee a look. Stars and feedback welcome!

Janee is open source under the MIT license. GitHub | npm