Every AI agent hosted on MyBotCloud can call a language model without you ever creating an account with OpenAI, Anthropic or any other provider. The platform holds its own OpenRouter key and proxies the call — that's the AI model proxy.
How it works
Three environment variables are injected automatically into every agent, on its first deploy:
| Variable | What it's for |
|---|---|
MYBOTCLOUD_AI_URL | the proxy's endpoint, which your code calls instead of going straight to a provider |
MYBOTCLOUD_AI_KEY | this app's own credential, write-only — never your account key, never reused by another app |
MYBOTCLOUD_AI_MODEL | the model picked when the app was created (you can change it later) |
Your code only needs one POST:
const res = await fetch(process.env.MYBOTCLOUD_AI_URL, {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.MYBOTCLOUD_AI_KEY}`,
"Content-Type": "application/json",
},
body: JSON.stringify({
model: process.env.MYBOTCLOUD_AI_MODEL, // optional — without it, the free default model is used
messages: [{ role: "user", content: "hi" }],
}),
});
const { content } = await res.json();
Which model to pick
The catalog is curated on purpose — this is our own OpenRouter balance that every agent spends from, so "every model that exists" would not be a safe default:
| Model | Cost | Plans |
|---|---|---|
| Nemotron Ultra | free | Start, Pulse, Orbit |
| Llama 3.3 70B | free | Start, Pulse, Orbit |
| GPT-4o mini | paid (our balance) | Pulse, Orbit |
| Claude 3.5 Haiku | paid (our balance) | Orbit |
The agent-creation screen already shows a picker with exactly which models your plan unlocks (locked ones are marked with the plan that unlocks them). To switch later without recreating the app, edit MYBOTCLOUD_AI_MODEL under Manage → Variables and redeploy.
Limits, on purpose
- 20 calls per minute per app — a floor against a runaway loop, until a configurable spend cap exists.
- Up to 20 messages per call, 8,000 characters each.
- Without
OPENROUTER_API_KEYconfigured on the server, the proxy answers503(turned off), never a generic network error.
Every call is logged (model, prompt/completion tokens, estimated cost) — today that's internal only, but it's the foundation for a future "runs" and cost tab, without you having to instrument anything in your own code.
Still stuck? Open a ticket in Support in the dashboard.