A router that makes your AI bills someone else's problem

What it does

Manifest sits between your application and 300+ LLMs across 17 providers, routing each request to whatever model makes sense. It accepts API keys, paid subscriptions you already own (ChatGPT Plus, Claude Pro, GitHub Copilot, even Kimi coding plans), and local models via Ollama or LM Studio — all through a single /auto endpoint. It also tracks spend, sets limits, and falls back to another model when one fails.

The interesting bit

The subscription reuse is the quietly clever part. Most gateways treat “bring your own key” as API keys only; Manifest also piggybacks on consumer and coding-tier subscriptions you might already be paying for, effectively turning them into infrastructure. The routing claims to weigh complexity and specificity, though the README doesn’t detail how those judgments get made.

Key highlights

• Single /auto endpoint covers OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, Qwen, Moonshot, MiniMax, Zhipu, BytePlus, OpenCode, OpenRouter, and custom OpenAI/Anthropic-compatible providers
• Local model support via Ollama (port 11434), LM Studio (port 1234), and llama.cpp (port 8080)
• GitHub Copilot integration uses OAuth, no API key required
• Cost tracking with notifications and spend limits
• Automatic fallback when a query fails
• Self-hosted Docker image or managed cloud version at app.manifest.build

Caveats

• Self-hosted: the first account created becomes admin, which is a footgun in shared environments
• Still in beta; the legacy npm package is deprecated and no longer maintained
• README claims “up to 70% cost savings” without showing methodology or benchmarks

Verdict

Worth a look if you’re running multiple AI providers and tired of wiring them up separately, or if you’re sitting on unused subscription tiers. Less compelling if you already have a clean routing layer you trust.