routerhubSI/uncensored-ai-api

RouterHub is not a library you install; it is a paid API gateway. If you already have code against the OpenAI or Anthropic SDK and want to switch model families, or try reduced-refusal variants without managing separate accounts, this is a low-friction route. If you need self-hosted inference or guaranteed model behavior, it is not.

The problem

A developer building against several model families usually maintains a key, billing relationship and SDK configuration per provider, and rewrites request code when switching. When the goal is to test a less-restrictive variant, the work grows: you must find a provider serving that variant and reconcile its interface. RouterHub collapses that into one account and one base URL, so switching models becomes a string change in the request. The result is faster experimentation across families, at the cost of moving model access to a third party.

What's different

The core mechanism is protocol compatibility rather than new tooling: requests follow the OpenAI chat-completions shape or the Anthropic messages shape, so existing clients that accept a custom base URL can point at the gateway. The README names Claude Code, Cursor, Codex CLI and OpenCode among compatible clients, and documents both SDK paths plus cURL. The genuinely distinguishing part is the catalogue: alongside mainstream families, RouterHub lists uncensored and abliterated variants "where available", which is the kind of access that is otherwise awkward to assemble.

The trade-off is twofold. First, availability is explicitly partial — not every family has a variant, and the README itself warns that reduced refusals are "not a guarantee" and recommends evaluating the model before deployment. Second, this is a gateway, not a model owner: behavior and terms belong to the underlying providers, so the gateway can only route, not change policy. Pay-per-token billing with prompt caching and card/crypto top-ups is stated, but no rate limits, latency figures or uptime commitments appear in the supplied material, so those remain open questions for anything latency-sensitive.

Who should try it

A good fit is an individual developer or small team prototyping across providers, or someone who wants to test abliterated variants in an existing OpenAI- or Anthropic-style codebase without rewriting request logic. The compatibility angle makes it especially cheap to trial from tools already configured for a custom base URL. A poor fit is anyone needing offline or self-hosted inference, strict data-residency control, or contractual guarantees about model behavior — the gateway model puts a third party in the request path by design. Teams needing per-provider compliance agreements would also find the abstraction unhelpful.

Requirements and limitations

Setup requires an account on the hosted service, a topped-up balance since billing is pay-per-token, and copying the exact base URL shown on the dashboard — the README's placeholder base URL is not the real endpoint. Ongoing upkeep includes monitoring spend across models and re-checking which variants are actually available, since that is stated as conditional. The bigger constraints are external: model terms and behavior rest with the providers, and the supplied documentation does not state rate limits, latency or service guarantees. Treat the uncensored variants as candidate models to evaluate, not as a turnkey policy removal.

HotRepo's take

The repository was created 2026-10-24, and HotRepo recorded it at #24 on the NEW-REPOSITORY board (all languages) on 2026-10-25 with 308 total stars — a single observation that is not a daily growth measurement. That is not enough to judge whether interest persists. Before committing, verify which model IDs and uncensored variants are actually present in the dashboard, and whether the terms of the underlying providers suit your use case.

Project source and documentation