MCP & AI Agents

Use keyzero as an MCP server so AI agents can resolve secrets and make authenticated requests without seeing raw credentials

The Problem

AI agents (Claude Code, Cursor, custom LLM agents) need to call authenticated APIs -- but giving them raw API keys means those keys end up in context windows, logs, and conversation history.

How keyzero Solves It

keyzero runs as an MCP (Model Context Protocol) server via kz server start --mcp, exposing two tools:

resolve -- runs the full pipeline (JWT verification, policy evaluation, secret resolution) and returns secrets to the agent
fetch -- resolves the credential and makes the HTTP request on behalf of the agent, injecting the credential based on credential_location. The agent sees the response but never sees the raw secret

The fetch tool is the preferred approach: it keeps secrets out of the agent's context entirely.

The Two Tools

resolve

Runs the full pipeline and returns resolved secrets.

Parameters:

Parameter	Type	Required	Description
`jwt`	string	yes	Bearer token for authentication
`resource`	string	yes	Resource ref path to resolve (e.g., `secret/data/prod/db/password`)

Returns: JSON with results containing per-ref entries with allowed, policy, mode, and value fields.

fetch

Resolves a credential and makes an authenticated HTTP request. The raw secret is never returned to the agent.

Parameters:

Parameter	Type	Required	Description
`jwt`	string	yes	Bearer token for authentication
`resource`	string	yes	Resource ref path to resolve credentials from
`resolver`	string	yes	Backend name to use for resolution
`url`	string	yes	Target URL to fetch
`method`	string	no	HTTP method (default: `GET`)
`headers`	object	no	Additional headers as key-value pairs
`body`	string	no	Request body

Returns: JSON with status, headers, and body from the upstream response.

The credential is injected based on the resolver's credential_location config (e.g., header:Authorization:Bearer or header:X-API-Key:).

Setup

1. Start the MCP Server

kz server start --bundle ./bundle.yaml --mcp

The --mcp flag makes kz server start communicate over stdio using the MCP protocol instead of starting an HTTP server.

2. Configure Your AI Tool

Claude Code

Add to .mcp.json in your project root:

{
  "mcpServers": {
    "keyzero": {
      "command": "kz",
      "args": ["server", "start", "--bundle", "./bundle.yaml", "--mcp"]
    }
  }
}

Example Flow

Agent receives a task: "Check the status of the production API"
Agent calls fetch with:
- jwt: the agent's authentication token
- resource: api/key
- resolver: key
- url: https://api.example.com/status
keyzero verifies the JWT, checks policies, resolves the API key from the backend
keyzero injects the credential into the request header and makes the HTTP call
Agent receives the response body and status -- but never saw the API key

credential_location

The credential_location field on resolver configs controls where the credential is placed in outgoing requests:

Format	Example	Result
`header:<name>:<prefix>`	`header:Authorization:Bearer`	`Authorization: Bearer <secret>`
`header:<name>:`	`header:X-API-Key:`	`X-API-Key: <secret>`

If not specified, defaults to header:Authorization:Bearer.