[UNTESTED] [brianroberg/api-proxy] 5 endpoint(s)
Gemini
API Proxy/Gemini
2,422 characters
Status: [UNTESTED]
Endpoints found:
http://localhost:8000/gmail/v1/users/me/labels
http://localhost:8000/gmail/v1/users/me/messages?maxResults=10&q=is:unread
http://localhost:8000/gmail/v1/users/me/messages/18d5a1b2c3d4e5f6?format=metadata
http://localhost:8000/gmail/v1/users/me/threads/18d5a1b2c3d4e5f6?format=full
http://localhost:8000/gmail/v1/users/me/labels/INBOX
Source: https://github.com/brianroberg/api-proxy
# API Proxy
A proxy server that enforces capability restrictions between AI agents and backend APIs. The proxy provides fine-grained access control that OAuth scopes cannot—allowing specific operations while blocking others that would be dangerous in agent hands.
## Overview
The proxy currently supports **Gmail** and **Google Calendar** APIs:
- **Gmail**: Allows read operations, label modifications, and draft management but **blocks all email sending capabilities**. This is necessary because Gmail's OAuth scopes don't provide fine-grained control: the `gmail.modify` scope (required for label changes) also grants send permission. The proxy provides the missing capability boundary.
- **Calendar**: Allows full event management (create, read, update, delete). Every event mutation — create, update, patch, delete, and RSVP — requires human confirmation when confirmation is enabled (the default `--confirm-modify` mode); reads never do. Operations that would send invitations to attendees are additionally highlighted in the confirmation prompt.
## Architecture
```
AI Agent (untrusted)
│
│ HTTP requests with API key (no backend credentials)
▼
api-proxy (this server)
│
├──► Invalid/missing API key → 401 Unauthorized
│
├──► Blocked operations → 403 Forbidden (always)
│
├──► Allowed operations → [Human confirmation if enabled]
│ │
│ ├── Approved → Forward to backend API
│ ├── Rejected → 403 Forbidden
│ └── Expired (nobody responded) → 403 confirmation_expired
│
│ Backend APIs (with credentials)
▼
Backend Services (Gmail, Calendar, etc.)
```
The AI agent never receives backend API credentials. It only knows how to talk to this proxy using its API key. A human operator can optionally review and approve operati