MCP-Warden Local Credential Broker

AI agents never see your real secrets, only scoped callable credentials

Problem

Pasting GitHub PATs, Stripe keys, and DB passwords directly into Claude Code or Cursor MCP configs means one prompt injection or leaked chat can expose everything. Developers need a local layer that isolates secrets and gives agents only controlled callable capabilities.

Solution

A local-first credential broker: secrets live only in the OS keychain, users define scopes like github-read or stripe-charges, and agents call high-level tools through a local proxy without ever seeing the raw secret. Every call is auditable and revocable.

Users

Developers using MCP agents like Claude Code and Cursor who need to connect sensitive APIs

MVP

Store secrets in the OS keychain; define scoped presets; expose a local proxy endpoint for MCP configs; log calls; one-click revoke

Revenue

Open-source core plus team subscription (policy sync, audit export); or one-time desktop license

Why now

The MCP ecosystem is expanding fast while secret security remains wild; local-first tools need no cloud operations and suit a trust-building indie product