Unified Billing & Ledger

A token router shows you your LLM spend. Floe shows you the whole bill — telephony, speech, search, compute, and LLM — in one ledger, priced per call, capped by one policy set.

Why this exists

A voice agent's cost is never just tokens. One conversation pays for telephony to place the call, speech-to-text to hear the caller, an LLM to decide what to say, and text-to-speech to say it back. If four vendors bill four different ways, you can't answer the only question that matters: what did this call cost, and was it worth it?

Floe answers it. Every paid leg — x402 vendors and LLM alike — settles from the same balance and lands in the same ledger, tagged by agent, task, and vendor. x402 vendors go through POST /v1/proxy/fetch; LLM calls go through Floe's OpenAI-compatible gateway — keyless POST /v1/chat/completions (recommended; fully-qualified provider/model ids) or the BYOK metered proxy POST /v1/llm/chat/completions (your own provider key; model id with or without a provider/ prefix). One balance, one ledger, one policy set.

Route your LLM through Floe (BYOK metered proxy)

POST /v1/llm/chat/completions is Floe's BYOK metered proxy: you bring your own upstream provider key and Floe charges only a small routing fee on top of provider cost. It is OpenAI-compatible — same request shape as OpenAI Chat Completions — but distinct from the keyless gateway in two ways: it prices from Floe's maintained LiteLLM cost map and accepts the model id with or without a provider/ prefix (gpt-4o or openai/gpt-4o — Floe strips it), and it needs your provider key. Point it at Floe's host, authenticate with your Floe agent key, and send your upstream provider key in the X-Floe-Provider-Key header — required on this endpoint (used only to call upstream, then discarded, never persisted). To skip the header, store the key once (encrypted, per provider) and use the keyless /v1 gateway instead — stored-key auto-routing works there, not on /v1/llm. See per-request vs stored keys for the storage flow and the fixed BYOK service fee.

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-4o",
    "messages": [{ "role": "user", "content": "Hello from Floe" }]
  }'

Don't want to hold a provider key at all? Use the keyless gateway POST /v1/chat/completions with a fully-qualified provider/model id (openai/gpt-4o) — Floe fronts the upstream relationship and bills you per call. That's the recommended path for new work; see Floe Inference (Keyless LLM & Voice).

Now LLM tokens, speech, and search all count against the same budgets, appear in the same analytics, and feed the same per-agent cost history — right next to the x402 vendors you already pay through the proxy.

One task, one budget, every vendor

Tag every leg of a job with the X-Floe-Task-Id header — on the proxy call and on the LLM call. The task budget then caps the sum across all vendors — telephony + STT + LLM + TTS — not each vendor in isolation.

-H "X-Floe-Task-Id: call-8842"

A $0.50 cap on call-8842 is a ceiling for the whole conversation, no matter how the cost splits across the four vendors that served it. See Spend Controls for policy types and windows.

Keyless inference

You can go one step further and drop the vendor accounts entirely. With keyless inference, you call LLM and voice models with no vendor account at all — Floe holds the upstream relationship and bills you per call from the same balance. Same ledger, same tags, one fewer key to manage.

See Floe Inference (Keyless LLM & Voice).

Next

Put the ledger to work behind a real voice agent — telephony, STT, LLM, and TTS on one bill: The Voice Stack.