How Agentbot Built the
First MiMo-Native
Agent Platform

When Xiaomi announced the MiMo V2.5 series with a 99% price reduction and 82 billion token plans, we saw an opportunity to build something no one else had: a managed AI agent platform where inference costs essentially disappear.

Agentbot is the first platform to integrate MiMo as a direct upstream provider — not through OpenRouter, not through a proxy, but straight to token-plan-ams.xiaomimimo.com. Every agent deployed on Agentbot runs on MiMo V2.5 Pro by default. No per-token charges. No surprise bills.

The Problem: Inference Costs Were Killing Us

Like every AI startup, we started on OpenRouter. GPT-4o, Claude Sonnet, Llama — the usual suspects. Our users were paying £29/month for a managed agent, and we were eating the inference costs. At scale, the math didn't work. A single active agent could burn through £5-10/day in API costs.

We needed a model that was fast, smart, and cheap. MiMo V2.5 Pro checked every box.

The Integration: Direct, Not Proxied

Most platforms route through OpenRouter or Vercel AI Gateway. We ripped that out and connected directly to Xiaomi's inference endpoint:

// opengateway.ts — MiMo direct upstream
const MIMO_BASE_URL = 'https://token-plan-ams.xiaomimimo.com/v1'
const MIMO_API_KEY = process.env.MIMO_API_KEY

function resolveGatewayUpstreams(): UpstreamConfig[] {
  const upstreams: UpstreamConfig[] = []

  // Direct MiMo — first priority, zero middleman
  if (MIMO_API_KEY) {
    upstreams.push({
      baseUrl: MIMO_BASE_URL,
      apiKey: MIMO_API_KEY,
      provider: 'xiaomi-direct',
    })
  }

  // OpenRouter — fallback only
  const openRouterKey = process.env.OPENROUTER_API_KEY
  if (openRouterKey) {
    upstreams.push({
      baseUrl: 'https://openrouter.ai/api/v1',
      apiKey: openRouterKey,
      provider: 'openrouter',
    })
  }

  return upstreams
}

The gateway tries MiMo first. If it fails (rate limit, timeout), it falls back to OpenRouter. In practice, MiMo hasn't failed yet.

BYOK: Users Bring Their Own MiMo Subscription

Power users can buy their own MiMo Max Monthly Plan (82B credits, ~$20/month) and paste their API key in Settings → BYOK. The gateway validates the key live against MiMo's API, stores it, and routes all their agent's requests through their personal subscription. Zero platform cost. Full speed.

This is the first BYOK system for a Chinese AI model on a Western agent platform.

The Numbers

What This Means for Agent Platforms

MiMo V2.5 changes the economics of AI agents entirely. When inference costs drop 99%, you can offer agents at flat monthly rates instead of per-token billing. Users don't think about usage. They just use their agent.

The V2.5 series also brings multimodality (images + text), 256K-1M context windows, built-in reasoning, and TTS models included free. This isn't a budget model — it's a frontier model at budget prices.

Open Source

We've open-sourced our MiMo integration layer. If you're building an agent platform, a coding assistant, or any AI product that needs fast, cheap inference — you can use our integration code directly:

github.com/Eskyee/agentbot-opensource

What's Next

We're working with the MiMo team on deeper integration: native TTS for agent voices, ASR for voice commands, and multimodal analysis for image-heavy workflows. The goal is an agent that can see, speak, listen, and think — all through MiMo.

Agentbot is live at agentbot.sh. Every agent runs on MiMo V2.5 Pro. Every plan includes inference. No per-token charges.

Written by the Agentbot team. Powered by MiMo. Built on OpenClaw.