ReliAPI

A tiny, self-hosted reliability layer for any HTTP API or LLM.
Retries, circuit breaker, cache, idempotency, cost caps, streaming — in one Docker container.

---

What is ReliAPI?

ReliAPI sits between your application and external APIs, adding reliability features that prevent failures and control costs. Works with any HTTP API (REST, payment gateways, SaaS) and any LLM provider (OpenAI, Anthropic, Mistral).

One Docker container. One config file. One unified API.

---

Why ReliAPI?

• 🔑 First-class idempotency — request coalescing prevents duplicate charges
• 💸 Predictable costs — soft/hard budget caps prevent surprise LLM bills
• 🚀 Universal proxy — same reliability features for HTTP and LLM APIs
• 📦 Minimal & self-hosted — no SaaS lock-in, full control over your data

Unlike LLM-only gateways (LiteLLM, Portkey), ReliAPI handles both HTTP and LLM requests. Unlike feature-heavy platforms, ReliAPI stays minimal and focused on reliability.

---

Key Features

• 🔄 Retries — exponential backoff with jitter
• ⚡ Circuit breaker — automatic failure detection per target
• 💾 Cache — TTL-based caching for GET/HEAD and LLM responses
• 🔑 Idempotency — request coalescing prevents duplicate execution
• 💰 Budget caps — soft (throttle) and hard (reject) cost limits
• 📡 Streaming — Server-Sent Events (SSE) for LLM responses
• 📊 Observability — Prometheus metrics and structured JSON logging

See all features →

---

Quick Start

Docker

docker run -d \
  -p 8000:8000 \
  -e REDIS_URL=redis://localhost:6379/0 \
  -e OPENAI_API_KEY=sk-... \
  -v $(pwd)/config.yaml:/app/config.yaml \
  ghcr.io/kikuai-lab/reliapi:latest

Configuration

Create config.yaml:

targets:
  openai:
    base_url: https://api.openai.com/v1
    llm:
      provider: openai
      default_model: gpt-4o-mini
      soft_cost_cap_usd: 0.10
      hard_cost_cap_usd: 0.50
    cache:
      enabled: true
      ttl_s: 3600
    auth:
      type: bearer_env
      env_var: OPENAI_API_KEY

---

Examples

HTTP Proxy

import httpx

response = httpx.post(
    "http://localhost:8000/proxy/http",
    headers={"Idempotency-Key": "req-123"},
    json={
        "target": "my-api",
        "method": "GET",
        "path": "/users/123"
    }
)
result = response.json()
print(result["data"])  # Response from upstream
print(result["meta"]["cache_hit"])  # True if cached

LLM Proxy

import httpx

response = httpx.post(
    "http://localhost:8000/proxy/llm",
    headers={"Idempotency-Key": "chat-123"},
    json={
        "target": "openai",
        "messages": [{"role": "user", "content": "Hello!"}],
        "model": "gpt-4o-mini"
    }
)
result = response.json()
print(result["data"]["content"])  # LLM response
print(result["meta"]["cost_usd"])  # Estimated cost

More examples →

---

Capabilities

Feature	HTTP APIs	LLM APIs
Retries	✅	✅
Circuit breaker	✅	✅
Cache	✅	✅
Idempotency	✅	✅
Budget caps	❌	✅
Streaming	❌	✅ (OpenAI)
Fallback chains	❌	✅

---

Documentation

• 📖 Full Documentation — complete guides and examples
• 🔑 Idempotency — request coalescing deep-dive
• 💰 Budget Caps — cost control examples
• 📡 Streaming — SSE events and examples
• ⚡ Circuit Breaker — failure detection
• 📊 Observability — metrics, logs, Grafana
• 🏗️ Architecture — system design
• 🚀 Deployment — production setup
• 🏢 Multi-Tenant — enterprise features
• 📚 API Reference — endpoint details

---

Comparison

Feature	ReliAPI	LiteLLM	Portkey	Helicone
Self-hosted	✅	✅	✅	❌
HTTP + LLM	✅	❌	❌	❌
Idempotency	✅ First-class	❌	⚠️	❌
Budget caps	✅	⚠️	✅	✅
Minimal	✅	❌	❌	❌

Detailed comparison →

---

License

MIT License - see LICENSE file for details.

---

Links

👉 Live Demo
👉 Documentation
👉 Issue Tracker

---

ReliAPI — Reliability layer for HTTP and LLM calls. Simple, predictable, stable.

Made with ❤️ by KikuAI Lab