GitHub - skorotkiewicz/llm-rt: Small Ruby prototype for an OpenAI-compatible LLM proxy with a refillable token bucket

2 min read Original article ↗

Small Ruby prototype for an OpenAI-compatible LLM proxy with a refillable token bucket.

It uses only Ruby standard libraries: no gems, no Rack, no WEBrick.

Run

BASE_API_URL=http://192.168.0.124:8888/v1 \
BASE_API_KEY=1mmer \
BASE_MODEL=gemma4 \
ruby llm_proxy.rb

The proxy listens on 0.0.0.0:8899 by default.

For your local LLM at 192.168.0.124:8888, run the saved local setup:

That starts the Ruby proxy at http://127.0.0.1:8899/v1 and forwards to http://192.168.0.124:8888/v1.

The saved local curl check is:

Manual equivalent:

curl -sS -i -m 60 http://127.0.0.1:8899/v1/chat/completions \
  -H 'Authorization: Bearer user-a' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gemma4",
    "messages": [{"role": "user", "content": "Reply with exactly: proxy ok"}],
    "max_tokens": 16
  }'

Verified result through the proxy: the upstream replied with proxy ok and the proxy returned X-RateLimit-Remaining: 0 with the local test bucket.

Run the smoke test:

Token bucket settings

MAX_TOKENS=10                 # max saved tokens per user
REFILL_TOKENS=2               # tokens added each refill
REFILL_INTERVAL_SECONDS=300   # 5 minutes
REQUEST_TOKEN_COST=1          # cost per accepted completion request

Each bearer token gets its own bucket. Requests without a bearer token are bucketed by remote IP. Set PROXY_API_KEYS=key1,key2 if the proxy should reject unknown client keys.

When the bucket is empty, /v1/chat/completions and /v1/completions return a normal OpenAI-style assistant response:

limit reached, wait 5 min

Test request

curl http://localhost:8888/v1/chat/completions \
  -H 'Authorization: Bearer user-a' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "anything",
    "messages": [{"role": "user", "content": "hello"}]
  }'

Optional estimated token mode

By default, one completion request costs REQUEST_TOKEN_COST bucket tokens. To charge roughly by prompt size plus expected output:

TOKEN_COST_MODE=estimate RESPONSE_TOKEN_RESERVE=256 ruby llm_proxy.rb

This is only an approximation for the prototype.