OmniCache

OmniCache is a lightweight, local caching proxy for Anthropic (Claude), OpenAI (GPT), and Google (Gemini) APIs.

When developing with AI agents (like Claude Code, Cursor, Aider, or custom LLM scripts), repeated prompts, test runs, and static file queries frequently make duplicate upstream API calls. OmniCache sits between your client and upstream providers to intercept matching requests locally in <1ms, saving API costs and eliminating remote network latency.

Installation

pip install omnicache-proxy

Quickstart

1. Start the Proxy Server

omnicache

By default, the proxy runs on http://localhost:8000. You can change the port with --port:

omnicache --port 8080

2. Connect Your Client

Claude Code (Terminal CLI)

Set the Anthropic base URL environment variable before running claude:

export ANTHROPIC_BASE_URL="http://localhost:8000"
claude

Python (OpenAI SDK)

Route the base_url parameter to the local proxy:

from openai import OpenAI

client = OpenAI(
    api_key="your-api-key",
    base_url="http://localhost:8000/v1"
)

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "How do I reverse a linked list in Python?"}]
)
print(response.choices[0].message.content)

Cursor / VS Code / Other Tools

In your tool's model settings, set the API Base URL to http://localhost:8000/v1.

Key Features

• Two-Tier Cache Engine:
  • L1 Exact Match (Trie Hash / Redis): Sub-0.05ms lookup for identical payloads.
  • L2 Semantic Match (Multi-Table LSH & FAISS ANN Indexing): Matches semantically equivalent prompts using high-speed multi-table hyperplane locality-sensitive hashing or memory-compacted FAISS HNSW.
  • Pluggable Embedders: Instant zero-dependency 512-d FastHashEmbedder or dense 384-d ONNX embeddings (all-MiniLM-L6-v2).
• Deterministic Coding-Agent Tool Replay:
  • Caches idempotent agent tool outputs (read_file, git_diff, grep_search, list_dir) with cryptographic Git workspace state fingerprinting (commit_sha:dirty_status_hash).
  • File modifications or git status changes instantly invalidate stale tool results with zero false positives.
• Client Explainability Headers:
  • Transparent X-OmniCache-Decision (HIT | MISS | BYPASS), X-OmniCache-Reason, and X-OmniCache-Similarity response headers across all OpenAI and Anthropic routes without mutating JSON body structures.
• Horizontal Scaling & Redis Clustering:
  • Pluggable storage adapter architecture supporting both zero-dependency standalone mode and distributed multi-worker/multi-replica clusters.
  • Secondary tenant index sets (omnicache:tenant_l1:{org_id}) and True LRU eviction using Redis Sorted Sets (ZSET).
  • Atomic spend tracking (INCRBYFLOAT) and sliding-window Redis RPM rate limiting across all worker processes.
  • Synchronized cluster-wide Circuit Breaker & upstream model failover state.
• Asynchronous Write-Behind Persistence:
  • Micro-batched non-blocking worker queue writing to embedded SQLite WAL store off the critical path with zero latency impact.
  • Durable Virtual Key and spend budget ledger surviving process cold starts.
• Remote Authenticated MCP JSON-RPC 2.0 Transport:
  • Native /mcp and /v1/mcp endpoint enabling AI IDEs (Cursor, Claude Code, VS Code) to perform intent-gated caching, vector search, and cache invalidation over HTTP.
• Agent Stream Replayer: Emulates natural token-streaming for cached responses so interactive CLIs (like Claude Code) stream smoothly without terminal glitches.
• Request Coalescing (SingleFlight): Deduplicates concurrent in-flight requests for the same prompt, making only one upstream call.
• Built-in CLI Utilities:
  • omnicache doctor: Checks database state, port bindings, and embedder health.
  • omnicache benchmark: Measures P50, P95, and P99 cache lookup latencies on your machine.
  • omnicache stats: Prints total tokens and cost savings directly to the console.
• Observability:
  • Web Dashboard: http://localhost:8000/dashboard
  • Prometheus Metrics: http://localhost:8000/metrics
  • CSV Ledger Export: http://localhost:8000/v1/cache/export

Horizontal Multi-Worker Deployment

To run OmniCache with multiple worker processes or in a clustered container environment, simply provide REDIS_URL:

# Multi-worker deployment with Redis distributed state
REDIS_URL="redis://127.0.0.1:6379/0" uvicorn server.gateway:app --host 127.0.0.1 --port 8000 --workers 4

Configuration

OmniCache can be configured via command-line flags or environment variables (in your shell or a local .env file):

Running Tests

Run the test suite using pytest:

git clone https://github.com/13manmayarai-hash/omnicache-proxy.git
cd omnicache-proxy
pip install -e ".[test]"
pytest tests/ -v

Documentation

• API Reference
• Architecture Overview
• Troubleshooting & FAQ
• Contributing Guidelines

License

MIT License. See LICENSE for details.