Graduate to 100% coverage

On a hosted orchestrator (Vapi / Retell / Bland), Floe governs what the platform lets it touch: the LLM leg pre-call on Vapi and Retell (Bland has no self-serve custom LLM, so its LLM — like every other leg — is governed by post-call reconcile), and everything else by reconcile. That's real governance — and it has a ceiling. The ceiling disappears when the agent runs on a self-hosted open voice stack — Pipecat or LiveKit — with every leg on Floe rails: 100% of spend cap-enforced — request legs gated pre-call, duration-billed legs (streaming STT, phone) metered live and cut off mid-stream at the cap — one key, one ledger, one cap.

The honest comparison. On an orchestrator, Floe-alone on the LLM leg is ~a token router with a budget. The answer isn't to pretend otherwise — it's this page: the path to 100%, which no token router can offer.

Leg-by-leg migration

Leg	On the orchestrator	On Floe rails
LLM	custom-llm → Floe (already pre-call)	Same endpoint: https://credit-api.floelabs.xyz/v1 — nothing to change
STT	platform's vendor, reconciled	Streaming WS wss://credit-api.floelabs.xyz/v1/audio/transcriptions/stream?model=deepgram/nova-3&encoding=linear16&sample_rate=16000&language=en — binary PCM in, interim/final transcripts out, metered per audio-second
TTS	platform's vendor, reconciled	POST /v1/audio/speech (OpenAI-compatible; e.g. openai/tts-1) — metered per character
Telephony	platform's carrier, reconciled	Floe Phone: POST /v1/developer/agents/{agentId}/numbers + POST /v1/calls, metered per minute

Pipecat — three drop-in services

The pipecat-floe package is the fastest path (pip install pipecat-floe):

from pipecat_floe import FloeSTTService, FloeLLMService, FloeTTSService

stt = FloeSTTService()                              # streaming STT over Floe's WS
llm = FloeLLMService(model="openai/gpt-4o-mini")    # OpenAI-compatible
tts = FloeTTSService(model="openai/tts-1", voice="alloy")

One FLOE_API_KEY — a floe_… agent key, not a floe_live_… developer key — powers all three; a single session cap bounds the whole run. Working references: the livekit-voice-agent recipe (LLM + TTS swaps, STT note) and floe-phone-sales-agent (every leg incl. telephony — the full dogfood).

Coverage & cost calculator

Illustrative 3-minute outbound call, typical talk ratio (~40% caller / ~60% agent speech, ~1.5k LLM tokens/turn × 12 turns). Recompute for your profile with POST /v1/estimate — rates move; the structure is the point.

	Orchestrator + Floe LLM	Full path on Floe
LLM (gpt-4o-mini class)	~$0.008 — pre-call ✓	~$0.008 — pre-call ✓
STT	platform rate, reconciled ⟳	~$0.010 (1.2 caller min × $0.0077 × 1.05) — pre-call ✓
TTS	platform rate, reconciled ⟳	~$0.011 (tts-1 @ $15/1M chars, ~700 chars) — pre-call ✓
Telephony	platform rate, reconciled ⟳	~$0.042 (Floe Phone outbound @ $0.014/min) — pre-call ✓
Platform fee	e.g. Vapi $0.05/min → $0.15	$0 (Floe margin is inside the leg rates above)
Pre-call coverage	~10–60% (the LLM share)	100%
Governance at the cap	LLM refused; other legs die next call	Every leg refused this call

Two honest caveats: orchestrator bundles include their platform value (builder UX, tooling) that raw legs don't replicate; and per-leg rates drift — trust POST /v1/estimate and the pricing page over this table.

Sequence it

1. Today: point the orchestrator's LLM at Floe + connect Reconcile Mode — full visibility, LLM pre-call.

2. Watch the coverage score (dashboard, per agent): it names the reconciled legs and what moving each one is worth.

3. Migrate legs when ready — STT and TTS are one-service swaps in Pipecat/LiveKit; telephony last (number porting is the sticky part).