8000
Skip to content
@Floe-Labs

Floe Labs

The cost & margin ledger for voice AI. True per-call cost across every vendor, margin per client, budgets enforced. BYOK, Vapi, Retell, Bland, Pipecat, LiveKit

Floe Labs

Floe Labs

What every voice call actually costs — across every carrier, speech, and model vendor. So finance can price contracts on real numbers, protect gross margin, and stop guessing at the gap between what they signed and what they deployed. Inside Vapi, Retell & Bland, or fully BYOK.

Website · Docs · Dashboard · 𝕏 @FloeLabs

npm pypi license


Start free. $3 Welcome Credit — 300 API credits on signup, no card required. Your agent makes its first paid call in minutes. Get started →

The problem

Every voice call your agent makes spends money across a dozen vendors — the phone carrier, the speech-to-text, the language model, the voice, the tools. Those bills arrive separately, in different units, on different days, and your finance team spends weeks stitching them together to answer one question: what did this client, this campaign, this call actually cost us?

Floe costs every call the moment it ends, tags the spend to the client and campaign, and shows you your margin per contract — so you can price the next deal on actuals instead of a blended guess, and walk into your next raise with your unit economics under control. (A token router meters only the LLM slice — ~40% of the bill — and is blind to the other 60%; on BYOK, platform dashboards report provider costs as $0. Floe is the join across every vendor, plus enforcement where it's in the path. Gateway overhead: 38ms p50 / ~180ms p99 on live keyless production traffic.)

Quickstart — one voice turn, one budget

A single spoken turn spends across speech-to-text, an LLM, and text-to-speech. Same key, same task budget for all three — a token router only ever sees the middle step.

# 1 · Transcribe (Deepgram) — through the Floe proxy
curl -X POST https://credit-api.floelabs.xyz/v1/proxy/fetch \
  -H "Authorization: Bearer $FLOE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Floe-Task-Id: call-8842" \
  -d '{"url":"<DEEPGRAM_STT_ENDPOINT>","method":"POST","body":"{...audio...}"}'

# 2 · Reason (any LLM) — keyless: only your Floe key, same ledger
curl -X POST https://credit-api.floelabs.xyz/v1/chat/completions \
  -H "Authorization: Bearer $FLOE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Floe-Task-Id: call-8842" \
  -d '{"model":"openai/gpt-4o","messages":[{"role":"user","content":"Book Friday at 2pm."}]}'

# 3 · Speak (ElevenLabs) — through the Floe proxy
curl -X POST https://credit-api.floelabs.xyz/v1/proxy/fetch \
  -H "Authorization: Bearer $FLOE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Floe-Task-Id: call-8842" \
  -d '{"url":"<ELEVENLABS_TTS_ENDPOINT>","method":"POST","body":"{...text...}"}'

All three legs share X-Floe-Task-Id, so one budget caps the whole conversation — STT, LLM, and TTS together, on one ledger. Each response returns its cost in X-Floe-Payment-Amount (e.g. 0.0125).

MCP (zero install) — add to Claude Code / Cursor / Claude Desktop:

claude mcp add --transport http floe https://mcp.floelabs.xyz/mcp \
  --header "Authorization: Bearer YOUR_FLOE_KEY"

Prefer an SDK? npm install floe-agent or pip install floe-agentkit-actions — see Repos.

Three ways in

1 · Inside your platform (Vapi / Retell / Bland). Documented hooks, zero platform cooperation:

  • Custom-LLM slot → Floe — pre-call enforcement on ~60% of call cost. Recipes: vapi-custom-llm, retell-custom-llm. (Bland has no self-serve custom LLM — govern it with Reconcile Mode below.)
  • Custom voice & transcriber → Floe (early access) — metered STT/TTS in your provider dropdown. Behind a flag until we publish a media-path latency benchmark — we don't promote a leg in your audio path ahead of its numbers. Recipe: vapi-voice-metered.
  • End-of-call webhooks → Reconcile Mode — every call reconciled onto one ledger at call-end; cross your cap and the next call is denied. A runaway campaign dies at call N, not call 10,000.
  • Coverage Score, per agent, in the dashboard — % of spend enforceable pre-call vs reconciled vs dark, and which leg to move to raise it.

Pre-call where we're in the path. Circuit breaker everywhere else. Setup → · Graduate to 100% coverage →

2 · BYOK. Keep your vendor accounts and keys — Floe meters, joins, and caps on top. Route just the LLM leg by changing the base URL and key:

from openai import OpenAI

client = OpenAI(
    base_url="https://credit-api.floelabs.xyz/v1",  # was https://api.openai.com/v1
    api_key=os.environ["FLOE_API_KEY"],             # your Floe key — not an OpenAI key
)
# every call now bills to your Floe balance, under your spend caps

OpenAI-compatible, so it works from any SDK. → Add Floe to your existing pipeline

3 · Keyless. One Floe key, no per-vendor accounts, welcome credits. Fund by card; settlement is automatic. Best for prototypes and net-new agents.

Vendor Marketplace — 2,000+ vendor API services, one key

Category Services
Compute Venice AI · OpenAI · Anthropic · Google Gemini · z.ai · Kimi
Voice Deepgram · ElevenLabs · Venice AI · Floe Phone via Twilio (live)
Image Venice AI (generation, upscale, edit, background removal)
Web Firecrawl (search + scrape)
Search Exa · Parallel AI · Tavily
Browser Hyperbrowser · Browserbase · Anchor Browser
Agent tools AgentMail · Pinata · PostalForm

Browse the full directory →

Works with your framework

Framework Status How
Coinbase AgentKit GA Native — floeActionProvider
LangChain GA getLangChainTools adapter
Vercel AI SDK GA getVercelAITools adapter
Claude / Cursor GA floe-mcp-server
CrewAI Beta via MCP server
OpenAI Agents SDK Preview MCP fallback; native adapter in progress
Plain HTTP/REST GA anything that speaks HTTP

Repos

Repo What it does Install
agentkit-actions TypeScript SDK — spend controls, metered vendor calls, agent awareness npm install floe-agent
agentkit-actions-py Python SDK — full parity pip install floe-agentkit-actions
floe-guard Local budget guardrail — hard-stops a runaway agent before it overspends pip install floe-guard
floe-cookbook Runnable end-to-end agents, including voice git clone
floe-mcp-server MCP server for Claude, Cursor, any MCP agent Setup
eve-floe Reference voice agent built on Floe git clone

Voice recipes (in floe-cookbook): vapi-custom-llm · retell-custom-llm · vapi-voice-metered


Built by operators from Airwallex, Western Union, eBay, Kado, Transak. hello@floelabs.xyz

Pinned Loading

  1. floe-mcp-server floe-mcp-server Public

    TypeScript 4 1

  2. floe-cookbook floe-cookbook Public template

    Agents and demos for the Floe finops for voice builders, including 15 LLM model providers + 2,000+ vendor API services

    TypeScript 3 1

  3. floe-guard floe-guard Public

    The spend meter and budget gate for AI voice agents. Meters STT + TTS + LLM + telephony per call, out of the box (Pipecat, LiveKit — Python & TypeScript). Hard-stops the next turn before it crosses…

    Python 517 95

  4. agent-skills agent-skills Public

    Agent skills for Floe — metered, budget-capped multi-vendor API access (LLM, STT, TTS, telephony, x402) through one key. For Claude Code, Cursor, and any Agent Skills client.

    1

  5. floe-cli floe-cli Public

    TypeScript 1

  6. pipecat-floe pipecat-floe Public

    Floe services for Pipecat — one Floe key for the LLM, STT, and TTS legs of a voice agent, metered per call with pre-call spend caps.

    Python 1

Repositories

Showing 10 of 13 repositories

Top languages

Loading…

Most used topics

Loading…

0