Drop-in AI proxy that tracks usage, caches requests, and caps your bills.
import OpenAI from "openai"; const client = new OpenAI({apiKey: process.env.OPENAI_API_KEY,baseURL: "https://api.openai.com/v1",billed every call});import OpenAI from "openai"; const client = new OpenAI({apiKey: process.env.OPENAI_API_KEY,baseURL: "https://api.mirkelo.com/v1",$0.00 Cost (Cached in 12ms)});One line. No SDK swap, no rewrite — Mirkelo speaks the OpenAI API.
Trusted By Leading AI Innovators
- Synergy Labs
- NeuroCorp
- Aura AI
- Cognition Systems
- Vector Foundry
Experience the Speed: Live Latency Test
Send a prompt through the simulated proxy, then send the exact same one again. The second call never reaches a provider.
Simulated for the demo. Real cache hits return in single-digit to low double-digit milliseconds.
One proxy between your app and every model.
Point your existing SDK at Mirkelo and get observability, caching, and spend control without touching your business logic.
Powerful configuration. Zero complex dashboards.
Budgets, fallbacks, and cache policy live in one type-safe file next to your code — reviewed in a pull request like everything else.
export default defineMirkeloConfig({projects: {"main-app": {monthly_budget_usd: 50,action_on_exceed: "fallback",fallback_model: "claude-3-haiku"}},caching: {strategy: "exact-match",ttl: "7d"}});Model-agnostic from the first request.
One base URL in front of every provider you already use. Swap the model string — or let a budget rule swap it for you — without touching the client.
OpenAI
gpt-4o · o-series
Anthropic
claude-opus · sonnet
Mistral
large · codestral
Llama (Local)
ollama · vLLM
Same OpenAI-shaped payload for all four. Mirkelo translates the rest.
Zero-Retention.
Zero Plain-Text Keys.
Mirkelo sits in your request path, so it is built to hold as little of your data as physically possible.
- Keys live in RAM only
- Your provider key is held in memory for the lifetime of a single request, then dropped. It is never written to the database, never logged, and never rendered back to a dashboard.
- Prompts are never stored
- We persist metadata — model, token counts, latency, cost — and nothing else. Prompt and completion bodies leave memory the moment the response is streamed back to you.
- Self-hostable anywhere
- Run the same proxy inside your own VPC with one container. No egress to our infrastructure, no vendor in your data path.
- Open source and auditable
- Every byte of the request path is MIT licensed on GitHub. Read it, fork it, or pin your own build.
- 01request inkey parsed into memory
- 02upstreamkey forwarded, never logged
- 03responsetokens + cost recorded
- 04teardownkey gone. prompt gone.
“Mirkelo slashed our token costs by 40% and improved our API performance overnight. The cache and budgeting tools are game-changers.”
Start free. Pay when the traffic is real.
Every plan includes the full proxy — caching, telemetry, and zero-retention key handling. No per-seat pricing.
Self-hosting is free forever, on every tier.
The questions engineers ask first.
Still unsure? The whole request path is open source — read it, or open an issue.
Stop overpaying for AI tokens today.
Get Started for FreeTakes 2 minutes to integrate. No credit card required.