keyper

██╗  ██╗███████╗██╗   ██╗██████╗ ███████╗██████╗
██║ ██╔╝██╔════╝╚██╗ ██╔╝██╔══██╗██╔════╝██╔══██╗
█████╔╝ █████╗   ╚████╔╝ ██████╔╝█████╗  ██████╔╝
██╔═██╗ ██╔══╝    ╚██╔╝  ██╔═══╝ ██╔══╝  ██╔══██╗
██║  ██╗███████╗   ██║   ██║     ███████╗██║  ██║
╚═╝  ╚═╝╚══════╝   ╚═╝   ╚═╝     ╚══════╝╚═╝  ╚═╝

your keys, kept — encrypted, local, MCP-native

Stop pasting sk-... into your AI. keyper keeps your API keys and env vars encrypted at rest on your own machine and hands them to your AI agent on demand over the Model Context Protocol.

Why

Every time you paste an API key into a chat, it lands in a transcript, maybe a log, maybe a training set. keyper breaks that habit. Your secrets live in one AES-256-GCM–encrypted file that only unlocks with your OS keychain or a passphrase. Your AI asks keyper for what it needs — and for the common case (calling an API, running a deploy) the value never even enters the conversation.

• 🔐 Encrypted at rest — AES-256-GCM, per-secret nonce, secret-name bound as AAD.

• 🗝️ Unlocks your way — OS keychain (silent) or a scrypt-derived passphrase.

• 🤖 MCP-native — works with Claude Desktop, Claude Code, Cowork, or any MCP client.

• 🧾 Redacted execution — run_with_secrets uses a key without ever showing it.

• 🖥️ Local web UI — add and name keys with a click; nothing leaves 127.0.0.1.

• 📄 One auditable file — ~800 lines of Python, two real dependencies.

Related MCP server: enigmagent-mcp

Quickstart

git clone https://github.com/sharziki/keyper
cd keyper
bash quickstart.sh

That installs the deps, creates your vault, prints your MCP config line, and opens the web UI to add keys. Prefer to do it by hand? Read on.

Install

pip install cryptography "mcp[cli]" keyring
# keyring is only needed for keychain mode; skip it if you'll use --passphrase

Create a vault

python keyper.py init                 # keychain mode (master key in your OS keychain)
python keyper.py init --passphrase    # or derive the key from a passphrase (scrypt)

Add secrets

python keyper.py set OPENAI_API_KEY -d "personal key"   # prompts (not in shell history)
echo -n 'sk-...' | python keyper.py set STRIPE_KEY       # or pipe it in
python keyper.py import-env ./.env                       # or bulk-import a .env
python keyper.py list                                    # names only, no values

…or skip the terminal entirely:

python keyper.py ui        # opens http://127.0.0.1:8765/?token=…

add · name · describe · delete — all over localhost, token-gated, values encrypted the instant you save

Connect it to your AI

All clients run the same stdio command: python keyper.py serve. Use absolute paths.

claude mcp add keyper -- /usr/bin/python3 /abs/path/keyper.py serve
# passphrase mode:
claude mcp add keyper -e KEYPER_PASSPHRASE=your-passphrase -- /usr/bin/python3 /abs/path/keyper.py serve

{
  "mcpServers": {
    "keyper": {
      "command": "/usr/bin/python3",
      "args": ["/abs/path/keyper.py", "serve"],
      "env": { "KEYPER_PASSPHRASE": "your-passphrase" }
    }
  }
}

Drop the env block if you use keychain mode. Restart the app afterward.

Then ask your AI things like "use run_with_secrets with OPENAI_API_KEY to list my models."

Tools the AI gets

The one tradeoff: get_secret puts a plaintext value into the model's context, where it could be logged. Prefer run_with_secrets — it injects secrets as environment variables into a subprocess and redacts any occurrence from the output. Treat a get_secret result as you would the key itself.

How it protects your secrets

• AES-256-GCM authenticated encryption; a fresh random 96-bit nonce per secret.

• The secret's name is bound in as Additional Authenticated Data, so a ciphertext can't be silently moved onto another name without decryption failing.

• Master key: 32 random bytes in your OS keychain, or scrypt (n=2¹⁵) from a passphrase. The vault file alone is useless without it.

• Plaintext is never written to disk or to the audit log.

• Vault file and access.log are created 0600.

• The web UI binds to 127.0.0.1, requires a per-launch token, and rejects any request whose Host header isn't localhost (blocks DNS-rebinding from malicious sites).

Threat model, briefly

keyper protects secrets at rest and keeps them out of your shell history and out of chat when you use run_with_secrets. It does not defend against malware already running as your user, or against you asking the AI to get_secret and pasting the result somewhere public.

Configuration

CLI: init · set · get · list · rm · rotate · import-env · ui · serve — run keyper <cmd> -h.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. keyper is intentionally small and auditable; changes that touch crypto or the UI's auth must spell out their threat-model impact.

License

MIT © 2026 Sharvil Saxena