Nataris — OpenAI-compatible P2P inference, drop-in config for LibreChat

[Sharrmavishal]

Built this over three months with a two-person team. Figured this community would find it useful — you can add Nataris as a custom endpoint in LibreChat in about 2 minutes.

What it is: Nataris is a P2P inference marketplace. Android phones run open-weight models locally (Qwen 2.5 0.5B, Llama 3.2 1B) and serve API requests via a standard OpenAI-compatible endpoint. Phone owners get paid per token. You get inference without managing any servers.

LibreChat config

endpoints:
  custom:
    - name: "Nataris"
      apiKey: "${NATARIS_API_KEY}"
      baseURL: "https://api.nataris.ai/v1"
      models:
        default: ["nataris-fast", "nataris-balanced"]
      titleConvo: true
      titleModel: "nataris-fast"
      summarize: false
      forcePrompt: false
      modelDisplayLabel: "Nataris"

Add NATARIS_API_KEY to your .env. That's it.

One thing to know: keep streaming enabled. Device cold-starts can cause Cloudflare timeouts on non-streaming requests.

Where we are

21 provider devices on the network, 2,775 inference jobs completed, 350K+ tokens processed in closed beta. The Android app just went live on Google Play.

Works well for longer workflows, batch tasks, anything where 5–20s latency is fine. Not great for real-time. No 7B models yet — waiting on desktop providers joining the network.

No prompt logging. No content filtering. $5 free credits on signup, no card needed.

API: https://api.nataris.ai/v1
Docs: https://api.nataris.ai/docs
Provider app (earn by running models on your Android): https://play.google.com/store/apps/details?id=ai.nataris.app