Add secure credential handling for browser automation

[shawnrushefsky]

Summary

• Adds secure credential handling so AI models can use passwords/API keys without seeing actual values
• Credentials stored locally (env vars or ~/.mcproxy/credentials.json) and referenced by name
• All browser responses automatically scrubbed to filter out any credential values

New Tools

Tool	Description
browser_list_credentials	List credential names (never values)
browser_type_credential	Type credential into selector by name
browser_keyboard_type_credential	Type credential at focused element
browser_set_credential	Store a credential locally
browser_delete_credential	Remove a stored credential

How It Works

Model: browser_type_credential(session_id, '#password', 'github_password')
         ↓
MCP Server: Resolves 'github_password' → actual value from local store
         ↓
Browser Server: Types actual value into the form field
         ↓
Model: Gets success confirmation (never sees the actual password)

Security Features

1. Reference-Only Access: Model only sees credential names, never values
2. Response Scrubbing: All browser responses filtered to replace credential values with [CREDENTIAL:name]
3. Local Storage: Credentials stored on MCP server machine, never sent to AI providers
4. File Permissions: Credentials file created with 600 permissions

Test plan

• Set credential via env var MCPROXY_CREDENTIAL_TEST=secret
• Verify browser_list_credentials shows the credential name
• Verify browser_type_credential types into a form field
• Verify response scrubbing replaces credential values in HTML content

🤖 Generated with Claude Code