AI Agents & Skills

Claude Code skill

PushKey ships a companion Claude Code skill that auto-activates when your conversation mentions API keys, secrets, .env files, or key rotation. It guides Claude through the correct MCP tool call sequence without manual lookup.

Trigger phrases

The skill activates when you say things like:

• "what keys do I have"
• "add this key to my vault"
• "set up env for my new project"
• "my STRIPE_KEY needs rotating"
• "inject secrets into /path/to/app"

How it works

Claude detects the trigger, calls unlock_vault, then uses the appropriate tools (list_keys, get_key, inject_env, etc.) in sequence. Key values are returned in the tool result — they never appear in the chat transcript.

OpenAI Agents SDK

Use PushKey as an MCP server in any OpenAI-compatible agent:

from agents import Agent, MCPServerStdio

pushkey = MCPServerStdio(
    command="python",
    args=["/absolute/path/to/pushkey_mcp.py"]
)

agent = Agent(
    name="DevAgent",
    mcp_servers=[pushkey]
)

The agent can then call unlock_vault, list_keys, inject_env, and all other tools listed in the MCP Server reference.

Any MCP-compatible host

PushKey implements standard MCP stdio and SSE transports. Any host that supports MCP works — LangChain, LlamaIndex, Cursor, Windsurf, etc.

# stdio (default)
python pushkey_mcp.py

# SSE
python pushkey_mcp.py --port 8765

Security in agent context

• Key values never appear in conversation history or model context
• The MCP server holds a short-lived in-memory session; lock_vault() clears it immediately
• Agents that call get_key() receive the value in the tool result only — not in the assistant message
• The server logs all tool calls to the encrypted audit log