ai-wrap

minimal go proxy for gemini llm with redis caching, mongodb logging, cost blocking, and key rotation

features

• proxies gemini api (same request/response format)
• cost tracking via headers + mongodb logs
• redis cache with mongodb fallback (temp < 0.3)
• blocks requests exceeding max cost (402)
• api key rotation from csv
• admin ui for monitoring

quick start

docker-compose up
# api: http://localhost:8089
# admin: http://localhost:3000

usage

curl -X POST http://localhost:8089/v1beta/models/gemini-2.0-flash:generateContent \
  -H "Content-Type: application/json" \
  -d '{"contents": [{"parts": [{"text": "hello"}]}]}'

response headers: X-Cost-Total, X-Cache-Status, X-Key-Source

docker images

builds push to docker hub on main:

• gyvastis/ai-wrap-api
• gyvastis/ai-wrap-app

config

config.yaml - models and costs (per 1M tokens usd)

env vars: PORT, MONGO_URI, REDIS_URI, GEMINI_TIMEOUT

api keys

optional data/keys.csv for key rotation:

key,provider,active,working_models,checked_at
AIza...,gemini,true,models/gemini-2.5-flash|models/gemini-2.0-flash,2025-12-17T11:25:25Z

• only active=true keys are used
• keys sorted by model priority (newest models first)
• on failure, rotates through all available keys
• if csv missing/empty, uses GEMINI_API_KEY env var

see CLAUDE.md for dev workflow, know-how/*.md for implementation details