Manifest

Open-source LLM router and gateway that uses a single OpenAI-compatible endpoint to access subscription-based, pay-as-you-go, local models and custom providers, automatically fixes erroneous requests, and switches to backup models in case of failures.

AI Developer Tools#Open Source#Routing#Cost

Updated: 2026-10-26

What is Manifest?

Manifest's entry point is a frequently overlooked number: about 5% of LLM API calls in production environments fail, and most of these failures could have been avoided. It claims to be "the only LLM Gateway that autofixes your bad requests on the fly" and has a straightforward goal—"LLM calls that don't break." With a single OpenAI-compatible endpoint, users can access subscription-based, pay-as-you-go, on-premises models and custom providers. The official statement says it can connect AI agents to any provider with various frameworks in just a few seconds, and supports free migration to self-hosting at any time after starting for free. Designed for AI builders, startup teams, and platform users, it is fully open for collaboration on GitHub.

Who is it for?

Engineering teams that integrate with multiple model vendors at the same time and are troubled by occasional failure rates and opaque billing; developers who need to flexibly switch between local models and cloud models; and teams building AI products, platforms and vendor services (the official categories of AI builders, Teams & startups, Platforms & providers). For budget-sensitive scenarios, you can first use the free tier to run real traffic, accumulate cost data, and then decide the investment method.

Key features

Single-endpoint Routing

One OpenAI-compatible entry point connects to all models and providers; no code changes are needed when switching providers. The official outlines the usage process in three steps: Route → Fix → Observe.

Model Fallback (Fallback)

Automatically switches to a backup model when the main provider is unavailable or throws an error, reducing service outages. The official slogan is "Fallback when providers are down".

Autofix

Corrects malformed requests on the spot. This is its core capability that differentiates it from other gateways, eliminating a large number of avoidable failures.

Comprehensive Logging and Cost Visualization

Tracks every request body, token, and dollar, giving you a clear overview of costs and usage.

Self-hosting & BYOK

Can be self-deployed, supports Bring your own key, local models and custom providers; fully open source (Built in the open) and remains OpenAI-compatible.

How to use Manifest

First, go through the process of requesting the key technical path to establish a log and cost baseline with it.
then enable rollback and auto-repair strategies, and gradually expand the access scope.
Before going live, complete supplier key rotation and fault drills, and treat the gateway as production infrastructure.
small teams can start with the free tier to verify the effect before deciding whether to self-host.
personal projects can also use it as a unified model access layer and keep an eye on the bills while using it.

Things to know

The gateway is a single point of failure for traffic; you are responsible for its deployment and operation, and failover strategies must be rehearsed in advance.

Autofix can handle format errors but cannot fix logical errors in the model itself; don't treat "Autofix" as a panacea.

When self-hosting, ensure centralized management and secure storage of API keys from various providers to prevent credentials from being scattered across configuration files.

Full logs will record the request body; for sensitive data involved, it is necessary to verify the log retention and desensitization strategy.

The self-hosted version requires a certain amount of operational effort, and you need to keep up with the gateway's own updates and security patches in a timely manner.

The functional differences between the free tier and self-hosted version are subject to the official website; please verify before making a selection.

When the rate limit or quota of a single provider is exhausted, the fallback strategy may not fully cover the risk, and you need to plan in advance for alternative providers and routing order.