Start Building
Read https://dev-dashboard.floelabs.xyz/agents.md and set up Floe for this project.
Paste into Claude Code, Cursor, or any coding agent.Read the quickstart
Works with
Anything that speaks HTTP — every model, every vendor, one key.
Open source

Start with the guardrail. No account.

floe-guard hard-stops a runaway agent before it burns your bill. Local, framework-agnostic, MIT.

$pip install floe-guard
For production teams

Running voice agents on Vapi, Retell, Bland, LiveKit or Pipecat?

15 minutes with the founders.

Book a demo
For production teams

Running production agents on Vapi, Retell, Bland, LiveKit or Pipecat?

15 minutes with the founders.

Book a demo
The whole bill

One key covers every line on the invoice

One voice call · what it actually coststotal $0.0300
TelephonyTwilio
$0.0072
TranscriptionDeepgram
$0.0041
LLM / reasoningOpenAI
$0.0127
SpeechElevenLabs
$0.0038
SearchExa
$0.0022
A token router sees 42%
Just the LLM slice — $0.0127 of the call.
Floe sees 100%
Every vendor, every line — $0.0300.
LayerVendorsWhat's live
LLM / reasoningOpenAI, Anthropic, Venice AIPay per call, any model, via one Floe key
Speech (STT / TTS)Deepgram, ElevenLabsSame key, same ledger, sub-cent per call
TelephonyFloe Phone via TwilioOn the same ledger + spend caps**
Memory / data / searchHydraDB, Exa, TavilyMetered per call, allowlisted per agent

** Floe Phone spend is unified onto the same ledger with the same spend caps.

Vendor Marketplace · 36+ services · 2,000+ payable APIs

Every leg of the call. One key. One ledger.

A voice call spends across 7–20 vendors. Floe meters and caps all of them — keyless on the gateway catalog, or BYOK with your own provider keys. These are the verified services live today; anything not listed is one submission away.

Verified services · by call legas billed on the ledger
Models / LLMthe ~40–60% leg
Any OpenAI or Anthropic model (BYOK metered — no model lock)·Venice AI full catalog (keyless — open-source to frontier, incl. Kimi)·Sarvam AI (Indic chat, 22+ Indian languages)
Realtime / speech-to-speech
OpenAI GPT Realtime·Google Gemini Live·xAI Grok Voice·dTelecom·Amazon Nova 2 Soniccoming soon·LiveKitcoming soon
Speech-to-text
Deepgram·OpenAI Whisper & Transcribe·AssemblyAI·ElevenLabs Scribe v2·Cartesia Ink-Whisper·Speechmatics·Azure·xAI·Sarvam Saaras·Venice·dTelecom
Text-to-speech
ElevenLabs·OpenAI TTS-1·Cartesia·Deepgram Aura-2·Hume·Rime·Inworld·MiniMax·Azure·Amazon Polly·xAI·Google Cloud TTS·Sarvam Bulbul·Venice·ForgeMesh
Telephony
Floe Phone — US numbers, inbound & outbound, Twilio-powered, live on the ledger
Search & data
Exa·Parallel AI·Tavily·Firecrawl
Browser
Hyperbrowser·Browserbase·Anchor Browser
Memory & toolspre / post call
HydraDB (vector memory)·AgentMail·Pinata Cloud·PostalForm·Venice image generation
No model lock, ever.

The BYOK metered proxy prices any model OpenAI or Anthropic ships — pass the model id, bring your key, get metering and a server-side cap. If a provider sells it and Floe can price it, it's callable.

Voice in 22+ Indian languages.

Sarvam's Indic chat, STT (Saaras), and TTS (Bulbul) run on the same key and ledger as the rest of your stack — regional-language agents without a separate vendor relationship.

Over budget? 402, pre-call.

Every service here is metered against the same balance and policies. When the cap is hit, the call is refused before any upstream spend — that's the whole point.

"Token and TTS usage varies every call, so true per-call cost only exists by joining usage data across providers after the call. Floe gives us that in real time."

CEO

$5M ARR voice agency

1/4

"Floe solves one of the missing pieces that voice AI builders need to operate autonomously in the real world: multi vendor billing and secure spending governance."

Ratikesh

Voice AI Developer at Pharma Company

2/4

"Shoutout to [Floe] app because cost tracking is HARD"

Naomi Carrigan

DevRel / Community Lead at Deepgram

3/4

"The problem Floe's solving is real. We hear some version of it from every team managing multi-model, multi-vendor agent stacks."

Partnerships Team

Z.ai

4/4

Keyless LLM · one key

Every model your agent needs — no per-provider accounts

Point your OpenAI-compatible SDK at Floe, pass your Floe key, and call models pay-as-you-go.

floe_•••key
Open-weight · live & keyless
LlamaQwenDeepSeekMistralGemmaCommand-R
FAQ
Your questions, answered

No. BYOK is live today: keep your own vendor accounts and keys — Floe meters, joins, and caps on top of them, and your platform spend flows onto the same ledger via end-of-call webhooks. Keyless — Floe fronts the vendor, one key, welcome credits — is the fast lane for prototypes and net-new agents.

Most production teams run BYOK; most prototypes start keyless. Switch anytime.

A voice call touches 7–20 vendors billing separately, and token and TTS usage varies every conversation — so true per-call cost only exists by joining usage data across providers after each call.

Floe does that join in real time and enforces budgets on top: caps rejected before money moves where Floe is in the path, and a circuit breaker that denies the next call everywhere else. One ledger: per-call, per-agent, per-vendor, per-customer.

Token routers meter your LLM spend — ~40% of a voice call's bill. The other ~60% — telephony, STT, TTS, search, memory — is spread across vendors they never see.

Floe meters the entire bill, enforces caps across all of it, and sits as the governance layer above your whole stack, token routers included.

Yes — through the platforms' own documented extension points, with zero platform cooperation. Point the custom-LLM slot at Floe for pre-call enforcement on ~60% of call cost; connect the end-of-call webhook and every call lands on one ledger with a between-call circuit breaker.

On BYOK, these platforms report provider costs as $0 — Floe is how you see and cap that spend anyway.

No — by design. A dropped call is a worse failure than an expensive one. Enforcement is pre-call admission: where Floe is in the path (LLM, STT, TTS, Floe Phone), an over-budget call is denied before it starts. Where it isn't, Reconcile Mode meters the cost at call-end and the circuit breaker denies the next call.

Your coverage score shows exactly which of your spend sits in which bucket.

A per-agent number in your dashboard: the % of that agent's spend that is pre-call enforceable vs post-call reconciled vs dark — across every platform it runs on. No single platform can compute this about itself, because no platform sees spend outside its own runtime.

The score also tells you which leg to move to raise it.

One Floe key covers the stack, live today: LLMs (OpenAI, Anthropic, Venice AI), speech (Deepgram, ElevenLabs), telephony (Floe Phone via Twilio), and memory, data & search (HydraDB, Exa, Tavily) — part of a marketplace of 36+ services across 8 categories, 2,000+ payable APIs.

Open-weight models — Llama, Qwen, DeepSeek, Mistral, Gemma, Command-R — run keyless via the gateway, pay-as-you-go.

Pick your lane. On a platform: one config field (custom-LLM slot) plus a webhook — no code. BYOK: point your OpenAI-compatible SDK at Floe's base URL with your Floe key; drop-in SDKs for AgentKit, LangChain, CrewAI, Vercel AI SDK, OpenAI Agents, and Claude (MCP). Keyless: paste one line into Claude Code or Cursor and your agent provisions itself. Or hit the REST API from anything that speaks HTTP.

Full quickstart: Agent Quickstart →

Fund by Visa, Mastercard, ACH, Apple Pay, Google Pay, or local methods in 100+ countries; settlement is automatic.

Pricing: free to integrate — no subscription, no seat fees, no minimums. Floe earns 5% on volume routed through it; vendors you call charge their own rates. BYOK only free.

Where Floe is in the path: per-call and daily caps, vendor allowlists, and time-bound permissions are enforced before payment — an over-cap call gets a 402, not a bill. Everywhere else: Reconcile Mode catches the cost at call-end and the circuit breaker denies the next call, across every platform at once.

Plus a self-serve kill switch: pause any agent instantly, or let a policy auto-trip it. A runaway campaign dies at call N, not call 10,000.

Alex Christian (DataMynt co-founder; payments, treasury, and compliance risk at Airwallex, Western Union, eBay) and Shivam Chaturvedi (DataMynt co-founder; CTO at Kado.money, acquired; founding engineer at Transak). Both shipped payments, compliance, and billing infrastructure at scale before Floe.

Keyless: the full financial loop
Pay any vendor per call. Authorize every dollar. Bill it from one place.
Step 1 · Fund
Fund your account.
From a card or bank in 100+ countries. Visa, Mastercard, Apple Pay, Google Pay.
Visa · Mastercard · Apple/Google Pay · 100+ countries
// Card or bank in. Vendors paid out.

  IN  Agent Wallet    OUT
  ────             ────
  Cards           Virtual cards
  Bank transfer    Bank accounts
  Apple Pay       Local payouts
  Google Pay
Step 2 · Pay
Pay as you go any vendor + model API, per call
Send any URL through the Floe proxy. The agent just gets its response back; it never touches the payment layer.
2,000+ payable APIs* · ~50ms signing
// Agent sends the target URL - that's it
await fetch("https://credit-api.floelabs.xyz/v1/proxy/fetch", {
  method: "POST",
  headers: { "Authorization": `Bearer ${apiKey}` },
  body: JSON.stringify({
    url: "https://api.anthropic.com/v1/messages",
    method: "POST",
    body: payload
  })
});

// → Floe settles the vendor, debits your ledger,
//   returns the response. No wallet, no SDK.
// Response header: X-Floe-Payment-Amount
Step 3 · Authorize
Set spend authorization it can’t blow past
Hard limits per agent, per vendor, per task — optionally time-bound — enforced before the call, not on next month’s invoice.
Pre-call enforcement · per-agent · per-vendor · time-bound
// Authorize spend at the agent + vendor level
await fetch("https://credit-api.floelabs.xyz/v1/spend-controls", {
  method: "POST",
  body: JSON.stringify({
    agentId: "agt_research_01",
    dailyCapUsd: 50,
    perCallCapUsd: 2,
    vendors: ["anthropic", "pinecone"], // allowlist
    expiresAt: "2026-06-30T00:00:00Z"
  })
});

// Any call over the cap or off the allowlist is
// rejected BEFORE money moves - a 402, not a bill.
Step 4 · Bill
One ledger across every vendor
Every payment — across all your APIs and all your agents — lands on one itemized ledger. Per-call cost, per-vendor totals, per-agent and per-customer attribution.
Per-call · per-vendor · per-agent · per-customer
// Unified spend ledger, filterable by agent
const feed = await fetch(
  "https://credit-api.floelabs.xyz/v1/developer/activity?agentId=agt_research_01"
)

// Returns an itemized, unified stream:
// [
//   { vendor: "anthropic", call: "messages",
//     costUsd: 0.012, ts, agentId, customerId },
//   { vendor: "pinecone",  call: "query",
//     costUsd: 0.003, ts, agentId, customerId },
//   ...
// ]

// Roll up by vendor, agent, or customer to bill.
Vendor Marketplace · 36+ services · 2,000+ payable APIs

Every leg of the call. One key. One ledger.

These are the verified services live today; anything not listed is one submission away.

Verified services · by call legas billed on the ledger
Models / LLMthe ~40–60% leg
Any OpenAI or Anthropic model (BYOK metered — no model lock)·Venice AI full catalog (keyless — open-source to frontier, incl. Kimi)·Sarvam AI (Indic chat, 22+ Indian languages)
Realtime / speech-to-speech
OpenAI GPT Realtime·Google Gemini Live·xAI Grok Voice·dTelecom·Amazon Nova 2 Soniccoming soon·LiveKitcoming soon
Speech-to-text
Deepgram·OpenAI Whisper & Transcribe·AssemblyAI·ElevenLabs Scribe v2·Cartesia Ink-Whisper·Speechmatics·Azure·xAI·Sarvam Saaras·Venice·dTelecom
Text-to-speech
ElevenLabs·OpenAI TTS-1·Cartesia·Deepgram Aura-2·Hume·Rime·Inworld·MiniMax·Azure·Amazon Polly·xAI·Google Cloud TTS·Sarvam Bulbul·Venice·ForgeMesh
Telephony
Floe Phone — US numbers, inbound & outbound, Twilio-powered, live on the ledger
Search & data
Exa·Parallel AI·Tavily·Firecrawl
Browser
Hyperbrowser·Browserbase·Anchor Browser
Memory & toolspre / post call
HydraDB (vector memory)·AgentMail·Pinata Cloud·PostalForm·Venice image generation
Voice in 22+ Indian languages.

Sarvam's Indic chat, STT (Saaras), and TTS (Bulbul) run on the same key and ledger as the rest of your stack — regional-language agents without a separate vendor relationship.

"Token and TTS usage varies every call, so true per-call cost only exists by joining usage data across providers after the call. Floe gives us that in real time."

CEO

$5M ARR voice agency

1/4

"Floe solves one of the missing pieces that voice AI builders need to operate autonomously in the real world: multi vendor billing and secure spending governance."

Ratikesh

Voice AI Developer at Pharma Company

2/4

"Shoutout to [Floe] app because cost tracking is HARD"

Naomi Carrigan

DevRel / Community Lead at Deepgram

3/4

"The problem Floe's solving is real. We hear some version of it from every team managing multi-model, multi-vendor agent stacks."

Partnerships Team

Z.ai

4/4

"Token and TTS usage varies every call, so true per-call cost only exists by joining usage data across providers after the call. Floe gives us that in real time."

CEO

$5M ARR voice agency

1/4

"Floe solves one of the missing pieces that voice AI builders need to operate autonomously in the real world: multi vendor billing and secure spending governance."

Ratikesh

Voice AI Developer at Pharma Company

2/4

"Shoutout to [Floe] app because cost tracking is HARD"

Naomi Carrigan

DevRel / Community Lead at Deepgram

3/4

"The problem Floe's solving is real. We hear some version of it from every team managing multi-model, multi-vendor agent stacks."

Partnerships Team

Z.ai

4/4

Pricing

Free to integrate and try BYOK. 5% per keyless spend through Floe.

$0to integrate
No subscription
No minimums
No seat fees
Get Started
Meet the team

Who's behind Floe

Floe Labs founders
Alex Christian
Co-founded DataMynt. Former Airwallex, Western Union, eBay.
Shivam Chaturvedi
Co-founded DataMynt. Previously CTO Kado.money (Acquired), Founding engineer at Transak.

Book a demo with the founders
15 minutes. We'll map your vendor bill on the call.
FAQ
Your questions, answered

No. BYOK is live today: keep your own vendor accounts and keys — Floe meters, joins, and caps on top of them, and your platform spend flows onto the same ledger via end-of-call webhooks. Keyless — Floe fronts the vendor, one key, welcome credits — is the fast lane for prototypes and net-new agents.

Most production teams run BYOK; most prototypes start keyless. Switch anytime.

A voice call touches 7–20 vendors billing separately, and token and TTS usage varies every conversation — so true per-call cost only exists by joining usage data across providers after each call.

Floe does that join in real time and enforces budgets on top: caps rejected before money moves where Floe is in the path, and a circuit breaker that denies the next call everywhere else. One ledger: per-call, per-agent, per-vendor, per-customer.

Token routers meter your LLM spend — ~40% of a voice call's bill. The other ~60% — telephony, STT, TTS, search, memory — is spread across vendors they never see.

Floe meters the entire bill, enforces caps across all of it, and sits as the governance layer above your whole stack, token routers included.

Yes — through the platforms' own documented extension points, with zero platform cooperation. Point the custom-LLM slot at Floe for pre-call enforcement on ~60% of call cost; connect the end-of-call webhook and every call lands on one ledger with a between-call circuit breaker.

On BYOK, these platforms report provider costs as $0 — Floe is how you see and cap that spend anyway.

No — by design. A dropped call is a worse failure than an expensive one. Enforcement is pre-call admission: where Floe is in the path (LLM, STT, TTS, Floe Phone), an over-budget call is denied before it starts. Where it isn't, Reconcile Mode meters the cost at call-end and the circuit breaker denies the next call.

Your coverage score shows exactly which of your spend sits in which bucket.

A per-agent number in your dashboard: the % of that agent's spend that is pre-call enforceable vs post-call reconciled vs dark — across every platform it runs on. No single platform can compute this about itself, because no platform sees spend outside its own runtime.

The score also tells you which leg to move to raise it.

One Floe key covers the stack, live today: LLMs (OpenAI, Anthropic, Venice AI), speech (Deepgram, ElevenLabs), telephony (Floe Phone via Twilio), and memory, data & search (HydraDB, Exa, Tavily) — part of a marketplace of 36+ services across 8 categories, 2,000+ payable APIs.

Open-weight models — Llama, Qwen, DeepSeek, Mistral, Gemma, Command-R — run keyless via the gateway, pay-as-you-go.

Pick your lane. On a platform: one config field (custom-LLM slot) plus a webhook — no code. BYOK: point your OpenAI-compatible SDK at Floe's base URL with your Floe key; drop-in SDKs for AgentKit, LangChain, CrewAI, Vercel AI SDK, OpenAI Agents, and Claude (MCP). Keyless: paste one line into Claude Code or Cursor and your agent provisions itself. Or hit the REST API from anything that speaks HTTP.

Full quickstart: Agent Quickstart →

Fund by Visa, Mastercard, ACH, Apple Pay, Google Pay, or local methods in 100+ countries; settlement is automatic.

Pricing: free to integrate — no subscription, no seat fees, no minimums. Floe earns 5% on volume routed through it; vendors you call charge their own rates. BYOK only free.

Where Floe is in the path: per-call and daily caps, vendor allowlists, and time-bound permissions are enforced before payment — an over-cap call gets a 402, not a bill. Everywhere else: Reconcile Mode catches the cost at call-end and the circuit breaker denies the next call, across every platform at once.

Plus a self-serve kill switch: pause any agent instantly, or let a policy auto-trip it. A runaway campaign dies at call N, not call 10,000.

Alex Christian (DataMynt co-founder; payments, treasury, and compliance risk at Airwallex, Western Union, eBay) and Shivam Chaturvedi (DataMynt co-founder; CTO at Kado.money, acquired; founding engineer at Transak). Both shipped payments, compliance, and billing infrastructure at scale before Floe.

FAQ
Your questions, answered

No. BYOK is live today: keep your own vendor accounts and keys — Floe meters, joins, and caps on top of them, and your platform spend flows onto the same ledger via end-of-call webhooks. Keyless — Floe fronts the vendor, one key, welcome credits — is the fast lane for prototypes and net-new agents.

Most production teams run BYOK; most prototypes start keyless. Switch anytime.

A voice call touches 7–20 vendors billing separately, and token and TTS usage varies every conversation — so true per-call cost only exists by joining usage data across providers after each call.

Floe does that join in real time and enforces budgets on top: caps rejected before money moves where Floe is in the path, and a circuit breaker that denies the next call everywhere else. One ledger: per-call, per-agent, per-vendor, per-customer.

Token routers meter your LLM spend — ~40% of a voice call's bill. The other ~60% — telephony, STT, TTS, search, memory — is spread across vendors they never see.

Floe meters the entire bill, enforces caps across all of it, and sits as the governance layer above your whole stack, token routers included.

Yes — through the platforms' own documented extension points, with zero platform cooperation. Point the custom-LLM slot at Floe for pre-call enforcement on ~60% of call cost; connect the end-of-call webhook and every call lands on one ledger with a between-call circuit breaker.

On BYOK, these platforms report provider costs as $0 — Floe is how you see and cap that spend anyway.

No — by design. A dropped call is a worse failure than an expensive one. Enforcement is pre-call admission: where Floe is in the path (LLM, STT, TTS, Floe Phone), an over-budget call is denied before it starts. Where it isn't, Reconcile Mode meters the cost at call-end and the circuit breaker denies the next call.

Your coverage score shows exactly which of your spend sits in which bucket.

A per-agent number in your dashboard: the % of that agent's spend that is pre-call enforceable vs post-call reconciled vs dark — across every platform it runs on. No single platform can compute this about itself, because no platform sees spend outside its own runtime.

The score also tells you which leg to move to raise it.

One Floe key covers the stack, live today: LLMs (OpenAI, Anthropic, Venice AI), speech (Deepgram, ElevenLabs), telephony (Floe Phone via Twilio), and memory, data & search (HydraDB, Exa, Tavily) — part of a marketplace of 36+ services across 8 categories, 2,000+ payable APIs.

Open-weight models — Llama, Qwen, DeepSeek, Mistral, Gemma, Command-R — run keyless via the gateway, pay-as-you-go.

Pick your lane. On a platform: one config field (custom-LLM slot) plus a webhook — no code. BYOK: point your OpenAI-compatible SDK at Floe's base URL with your Floe key; drop-in SDKs for AgentKit, LangChain, CrewAI, Vercel AI SDK, OpenAI Agents, and Claude (MCP). Keyless: paste one line into Claude Code or Cursor and your agent provisions itself. Or hit the REST API from anything that speaks HTTP.

Full quickstart: Agent Quickstart →

Fund by Visa, Mastercard, ACH, Apple Pay, Google Pay, or local methods in 100+ countries; settlement is automatic.

Pricing: free to integrate — no subscription, no seat fees, no minimums. Floe earns 5% on volume routed through it; vendors you call charge their own rates. BYOK only free.

Where Floe is in the path: per-call and daily caps, vendor allowlists, and time-bound permissions are enforced before payment — an over-cap call gets a 402, not a bill. Everywhere else: Reconcile Mode catches the cost at call-end and the circuit breaker denies the next call, across every platform at once.

Plus a self-serve kill switch: pause any agent instantly, or let a policy auto-trip it. A runaway campaign dies at call N, not call 10,000.

Alex Christian (DataMynt co-founder; payments, treasury, and compliance risk at Airwallex, Western Union, eBay) and Shivam Chaturvedi (DataMynt co-founder; CTO at Kado.money, acquired; founding engineer at Transak). Both shipped payments, compliance, and billing infrastructure at scale before Floe.