How Anuma routes 1.2 million AI requests through one open-source gateway

TL;DR: Anuma has routed over 1.2 million AI requests across 30+ models through Bifrost, Maxim AI’s open-source gateway. It is how memory stays private and portable across models.

ZetaChain is building the private memory layer for AI: a user-owned foundation for memory, identity, permissions, payments, and agents that works across every model and app. Part of that promise is technical. If your memory is going to follow you from model to model, the infrastructure underneath has to treat every model as interchangeable and every conversation as yours.

This post is about the gateway underneath. Since February, Anuma has routed over 1,200,000 AI requests and more than 30 billion tokens through Bifrost, the open-source AI gateway from Maxim AI, across 30+ models and more than a dozen providers, from a single self-hosted deployment.

One gateway for every model

Anuma is a consumer AI app that remembers you across models. Users pick from frontier and open-weight models, or let Anuma choose for them, and their memory stays with them either way. That only works if switching models is a routing decision rather than a product rewrite. It is also why we are exploring opening that routing layer to developers, so other apps and agents can build on the same router and, with the user’s permission, the same memory. More on that below.

We chose Bifrost as Anuma’s gateway in November 2025, nine days into the codebase, before a single direct provider integration existed. Every model call Anuma has ever made, chat, reasoning, vision, embeddings, and tool execution, has gone through it. Bifrost is open source under Apache 2.0, runs in our own cluster, and speaks an OpenAI-compatible API, so our backend talks to one endpoint and Bifrost handles the differences between providers.

How a request moves through Anuma

Every Anuma model has a stable public name. Behind it, our router keeps an ordered list of providers that can serve it. Bifrost receives the primary and the fallbacks in the same request, so if the first provider is slow or down, the request moves down the list without the user noticing. Today that map covers more than 50 model identities served by 14 providers, including Fireworks, OpenAI, Anthropic, OpenRouter, Cerebras, xAI, DeepInfra, Google, and Alibaba, plus an open-weight embedding model, Qwen3-Embedding-8B, that we host ourselves and reach through the same gateway.

When a user lets Anuma choose the model, a classifier sorts the request into general, code, reasoning, vision, or image work and picks from a pool that matches the user’s privacy setting. Private mode only draws from open-weight models.

Tool calls take the same path. Anuma’s web search, page reading, vision, image, video, and market data tools are Model Context Protocol servers, and Bifrost executes them on the model’s behalf. Pricing comes from Bifrost’s model catalog too, so every request is settled against a known cost before the user’s credits are touched.

Private by design

Anuma’s privacy model starts in the app. Every account is a self-custodial wallet provisioned by Privy, a Stripe company, and a wallet-derived key is the user’s AI identity. Memory is encrypted on the user’s device and only that key opens it. We covered the identity layer in our Privy announcement. The gateway layer is where that promise is enforced on the way to the model:

• Conversations are never stored on our servers. The backend records which model served a request, how many tokens it used, and what it cost. It does not keep the content.

• Private mode means open-weight models only. The routing rules that enforce that live in one place, so a new model or provider inherits them automatically.

• Private image generation is enforced server-side. If a request is marked private, the router rewrites the image model to a private one regardless of what the model asked for.

• Telemetry carries no content. We track latency, cost, and provider health per request, never prompts or outputs.

1.2 million requests and counting

Anuma launched publicly this spring. Since February, Anuma has sent over 1.2 million inference requests through Bifrost, and more than 3 million gateway calls once embeddings are counted, with a failure rate well under one tenth of a percent at the gateway. Those requests come from nearly 300,000 wallets created on Anuma, each one a self-custodial AI identity. The live numbers, including wallet growth, weekly request volume, and the model mix, are published openly on ZetaChain Research.

What comes next

Bifrost exposes detailed Prometheus metrics for every provider it talks to. We are building on those to rank providers dynamically per model, using first-token latency and success rate over a rolling window, so the fallback order adapts to how each provider is performing right now. That work is in active development and on the roadmap for Anuma. Alongside it, we plan to keep expanding the open-weight pool that private mode draws from.

We are also exploring opening the router itself to developers. The same routing layer that serves Anuma, with its provider fallbacks and privacy pools, could sit behind an API that other apps and agents call, alongside shared memory that users choose to grant access to. Bifrost would remain the gateway underneath, so every request from a third-party app would carry the same failover and cost accounting Anuma gets today. We will share more as that work takes shape.

Try it

Anuma is live at anuma.ai. Developers can find Bifrost on GitHub and get started with the gateway quickstart.

About ZetaChain

ZetaChain is the foundation for private AI, with user-owned private memory, private inference, identity, permissions, payments, and agents across connected models and apps. It powers Anuma, the first consumer AI product built on ZetaChain: a private AI app that remembers, combining leading models, encrypted memory, and user-controlled access in one experience.

Follow ZetaChain on X (Twitter) and join the conversation on Discord and Telegram.

Categories