Compute

LLM inference and embeddings — payable with Floe credit on Base.

Two ways to buy compute through the spend layer:

• Floe Metered LLM (native) — one OpenAI-compatible endpoint that fronts any OpenAI or Anthropic model. No model lock. Billed at-cost per token to your credit line, capped server-side. (Below.)

• x402 compute vendors — third-party providers (e.g. Venice AI) reached through the x402 proxy. (Further down.)

---

Floe Metered LLM (native) — no model lock

Chat: POST https://credit-api.floelabs.xyz/v1/llm/chat/completions · OpenAI-compatible Embeddings: POST https://credit-api.floelabs.xyz/v1/llm/embeddings Price: at-cost per token (+5% buffer) · debited from your Floe credit line · capped server-side

Call any model OpenAI or Anthropic ships — pass the model id in the request body. There is no model allowlist and no model lock: if a provider sells it and Floe can price it, the proxy serves it. The featured models below are just current favorites, not a fixed menu.

Two headers do the work:

Header	Value	Purpose
Authorization	Bearer floe_<agent key>	Authenticates the agent and bills its credit line
X-Floe-Provider-Key	<your OpenAI/Anthropic key>	Pass-through key used to call upstream — never stored

This BYOK proxy prices from Floe's maintained LiteLLM cost map — there's no model allowlist; pass any model your provider serves. (The keyless gateway's catalog is separate — see Keyless Inference or GET /v1/models.) The exact USDC charge for each call is returned in the X-Floe-Cost-USDC response header.

This /v1/llm surface is the BYOK metered proxy — you bring a provider key; Floe prices the call from its LiteLLM cost map and accepts the model id with or without a provider/ prefix (gpt-5.5 or openai/gpt-5.5 — the prefix is stripped). The keyless gateway at /v1/chat/completions is different: no provider key, and it resolves an exact catalog id — the fully-qualified provider/model (a bare gpt-5.5 is rejected). See Keyless Inference.

Drop-in OpenAI client

from openai import OpenAI

client = OpenAI(
    base_url="https://credit-api.floelabs.xyz/v1/llm",
    api_key="floe_<agent key>",                         # -> Authorization: Bearer ...
    default_headers={"X-Floe-Provider-Key": "sk-..."},  # your OpenAI/Anthropic key
)

resp = client.chat.completions.create(
    model="claude-sonnet-4-6",   # any OpenAI/Anthropic model — no lock
    messages=[{"role": "user", "content": "Hello from Floe"}],
)

curl -X POST https://credit-api.floelabs.xyz/v1/llm/chat/completions \
  -H "Authorization: Bearer $FLOE_API_KEY" \
  -H "X-Floe-Provider-Key: $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.5","messages":[{"role":"user","content":"Hello from Floe"}]}'

When the credit line (or session cap) is exhausted, the call is refused with 402 before any upstream request runs — spend stops at the cap.

---

x402 compute vendors

Service	Endpoints	Price	Status
Venice AI	Chat Completions, Responses, Embeddings	metered / tokens	Verified
Sarvam AI	Chat Completions (Indic LLM)	metered / tokens	Verified

Venice chat, responses, and embeddings are served through Floe's drop-in OpenAI-compatible surface — point any OpenAI SDK at https://credit-api.floelabs.xyz/v1/venice with your Floe agent key. No Venice key, no request envelope. See the full guide below: Venice models.

Sarvam AI chat is served on Floe's OpenAI-compatible gateway — call POST /v1/chat/completions with the model sarvam/<id> and your Floe agent key. No Sarvam key. Its proprietary voice, translation, transliteration, language-ID, and document models run through the marketplace shim (see Voice). See the full guide below: Sarvam AI — Chat Completions.

---

Sarvam AI — Chat Completions

Endpoint: POST https://credit-api.floelabs.xyz/v1/chat/completions · model sarvam/<id> Price: metered per token · billed from your Floe balance

Indic-language chat (Sarvam 105B / 30B) across 22+ Indian languages — India's sovereign-AI stack. OpenAI-compatible; set stream: false. Prices are Sarvam's INR list converted at ~₹83/$ plus a 5% Floe margin.

See Sarvam AI for the full model list and per-token pricing.

curl -X POST https://credit-api.floelabs.xyz/v1/chat/completions \
  -H "Authorization: Bearer $FLOE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"sarvam/sarvam-105b","messages":[{"role":"user","content":"नमस्ते! भारत की राजधानी क्या है?"}]}'

---

Venice models

Venice exposes its full chat and embedding catalog through Floe's metered Venice routes (separate chat, responses, and embeddings endpoints, below) — from fast open-source to frontier, including private TEE-attested (E2EE) and uncensored options. Pass the model id straight through; there's no allowlist. The sections below cover how to call them and how to pick by capability; the full catalog with per-token pricing is served from the live endpoint below.

Discover it live (capabilities, context windows, pricing) from Venice's public, no-key endpoint:

curl "https://api.venice.ai/api/v1/models?type=text"        # chat / reasoning models
curl "https://api.venice.ai/api/v1/models?type=embedding"   # embedding models

Capabilities shown below: reasoning · vision · code · audio in · video in · E2EE · X search. Most models additionally support tool/function calling, structured output, and web search. Prices are USD per 1M tokens, track Venice's catalog (refreshed ~weekly), and may change — you're billed the exact metered amount in X-Floe-Cost-USDC.

The full current catalog — every chat and embedding model with per-token pricing — is served from the live, no-key endpoint above and mirrored on the marketplace card /vendors/venice-compute, the canonical Venice model reference. Pass any listed id as the model field.

---

Venice AI — Chat Completions

Endpoint: POST https://credit-api.floelabs.xyz/v1/venice/chat/completions Price: metered per token · billed from your Floe balance

OpenAI-compatible chat across text & reasoning models, with tool use. Set stream: false.

curl -X POST https://credit-api.floelabs.xyz/v1/venice/chat/completions \
  -H "Authorization: Bearer $FLOE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"kimi-k2-5","messages":[{"role":"user","content":"Hello from x402"}]}'

Venice AI — Responses

Endpoint: POST https://credit-api.floelabs.xyz/v1/venice/responses Price: metered per token · billed from your Floe balance

Venice's Responses API — structured output and tool calls with built-in state.

curl -X POST https://credit-api.floelabs.xyz/v1/venice/responses \
  -H "Authorization: Bearer $FLOE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"kimi-k2-5","input":"Explain x402 in one sentence."}'

Venice AI — Embeddings

Endpoint: POST https://credit-api.floelabs.xyz/v1/venice/embeddings Price: metered per token · billed from your Floe balance

Vector embeddings for retrieval, clustering, and semantic search.

curl -X POST https://credit-api.floelabs.xyz/v1/venice/embeddings \
  -H "Authorization: Bearer $FLOE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"text-embedding-bge-m3","input":"x402 makes agents pay per call"}'

Pick a model

The curated Venice chat models — with model ids, latency tier, context window, and tool-calling support — live on the marketplace at /vendors/venice-compute. That card is the canonical model reference; pass any listed id as the model field. For a voice agent, start at the fast end (e.g. a small instruct model) and reach for a frontier reasoning model only when reasoning matters more than turn speed.

---

Voice agents: Vapi + Venice through Floe

A voice agent has two cost centers: the model's tokens (its brain) and the paid tool calls it makes (search, transcription, your own APIs). Route both through the Floe proxy and a single credit line meters and caps both — one ceiling, no separate Venice key or USDC to manage.

The reference implementation is vapi-floe-poc. It works like this:

1. Point your Vapi assistant's LLM at a custom-llm provider whose URL is a tiny shim you host.

2. The shim forwards Vapi's OpenAI-shaped chat request to Venice's chat endpoint through the Floe proxy (POST /v1/proxy/fetch → api.venice.ai/api/v1/chat/completions).

3. The agent's tools already call paid APIs through the same proxy with the same Floe key.

Because every request — model and tools — debits the same Floe credit line, your session cap is a unified ceiling. When it's exhausted, the next call is refused with 402 before any upstream spend.

The shim (the "compute" plane) forwards one request:

// Vapi POSTs an OpenAI chat request here; we pay Venice through Floe.
const veniceBody = JSON.stringify({ model, stream: false, messages, tools });

const res = await fetch("https://credit-api.floelabs.xyz/v1/proxy/fetch", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.FLOE_API_KEY}`,  // same key as tools
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    url: "https://api.venice.ai/api/v1/chat/completions",
    method: "POST",
    headers: { "Content-Type": "application/json" },
    body: veniceBody,  // Venice request, serialized to a string
  }),
});

const completion = await res.json();
const costUsdc = res.headers.get("X-Floe-Payment-Amount");  // per-call cost

The Vapi assistant points its model at that shim:

model: {
  provider: "custom-llm",
  url: `${SERVER_URL}/llm`,   // your shim — forwards to Venice via Floe
  model: "llama-3.2-3b",      // any id from /vendors/venice-compute
}

Pick the model id from the marketplace card at /vendors/venice-compute — for voice, a small low-latency model with tool calling keeps turns snappy. Venice's raw x402 settlement (Sign-In-With-X, balance, top-up, payment headers) is entirely Floe's job; the shim only needs a Floe API key.