Mantis

Mantis

2 min read Original article ↗

SELF_HOSTED: TRUECLOUD_PROVIDER: AWSAPI_STYLE: OPENAI

A self-hosted LLM gateway for routing, caching, guardrails, and observability across model providers.

01 // UNIFIED_API

One Interface

A stable chat completions endpoint that sits in front of multiple model targets. Switch providers without changing a single line of client code.

OPENAICLAUDE+MORE

02 // INFRASTRUCTURE

AWS-Native

Built for the AWS ecosystem. Deploys via Terraform to ECS, using ElastiCache for fast caching and Bedrock for foundational models.

ECSCLOUDWATCHBEDROCK

03 // LOGIC_ENGINE

Policy Driven

Routing, retry, and fallback logic. Define timeouts, cooldown periods, and cache behavior directly in your config.

ROUTINGRETRYCACHING

// Capabilities

01

Configurable Routing

Route requests by metadata, model aliases, weighted targets, and fallback chains.

multi-providerweighted routingfallback chains

Configurable routing: stepped medallion with a pulsing core node

02

Gateway Orchestration

Coordinate validation, cache checks, cooldowns, provider calls, retries, and terminal responses.

lifecycle controlshort circuitretry logic

Gateway orchestration: square wave field in a diamond aperture with one packet on the centre line

03

OpenAI-style API Surface

Send chat completion requests through a single gateway endpoint with optional routing metadata.

streaming supportunified schemasingle endpoint

OpenAI style API surface: split circle with paired registers and a packet on the axis

04

Python SDK

Call Mantis from application code without manually constructing each HTTP request.

pip installableasync clientrouting metadata

Python SDK: call grid with one active cell, routed through by a packet

// Team

Hubert de Parseval headshot

Hubert de Parseval Europe

Rey van den Berg headshot

Rey van den Berg Dublin, Ireland

Riz Syed headshot

Riz Syed London, UK

Samuel Meka headshot

Samuel Meka Taipei, Taiwan

// Next Steps

Route. Cache. Guard. Observe.