One key for every live model. Credit sales open at launch.
Network Agent
Developer guide

Build on every
supported model.

One OpenAI-compatible endpoint, one key and one balance. The API opens at launch; this guide is what it will accept.

What it does

A named model is always the model that answers — it is never swapped for another. Ask for agent/auto or a profile instead, and the agent picks a live model that can handle the request, then names it in the response.

Base URL

Use this with any OpenAI-compatible client:

https://api.networkagent.site/v1

Authenticate with Authorization: Bearer YOUR_KEY.

Your first key

  1. Connect a wallet on Robinhood Chain to open your workspace.
  2. Top up with card, USDC, ETH or $NETWORKAGENT. Every top-up is doubled.
  3. Create a key. It is shown once; store it on your server, never in browser code or a public repository.
  4. Send a request to any live model.

Choosing a model

Call GET /models to list the exact IDs your key can use. Pass one as model to pin that model, or pass a profile to let the agent choose.

Agent Auto

Set model to agent/auto, or to agent/<profile>. The optional routing_profile field accepts auto, balanced, cheap, fast, best, reasoning, code, vision.

The agent drops models that cannot meet the request (context length, tools, images, JSON output), ranks the rest by the profile, and sends it to the first. If that model refuses before producing output, the next one is tried and the switch is reported in the x-agent-fallback header. Once output has started, it never switches.

request
curl https://api.networkagent.site/v1/chat/completions \
  -H "Authorization: Bearer $NETWORK_AGENT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agent/auto",
    "routing_profile": "balanced",
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

Endpoints

GET/modelsEvery model your key can call, with prices and limits.
POST/chat/completionsChat requests, streamed or whole. OpenAI request and response shape.
POST/embeddingsVector embeddings from any embedding model in the catalogue.
GET/balanceDollar credit, holder tokens and usage so far for the key's workspace.

Balance and billing

Each request is charged at the answering model’s listed rate for input, cached, reasoning and output tokens. Holder tokens are spent first, then dollar credit. GET /balance shows both.

Streaming

Set stream: true for server-sent events in the OpenAI format. The final event carries token counts and the cost of the request.

Errors

  • 401 — the key is missing, wrong or revoked.
  • 402 — the workspace balance cannot cover the request.
  • 404 — no live model matches that ID.
  • 429 — slow down; the key hit its rate limit.
  • 503 — no model for that profile is available right now.