Reference AI agents built on Floe — the spend layer for agents. One Floe key pays every vendor, with programmable spend controls.
Website · Dashboard · Docs · Quickstart
One key for your agent's whole vendor bill — LLM, voice, telephony, search, data — metered per call and budget-capped. Let your coding agent set it up, or wire it yourself:
| Path | One line |
|---|---|
| Agent — Claude Code / Cursor does the setup | paste: Read https://dev-dashboard.floelabs.xyz/agents.md and set up Floe for this project. |
| Skill — install the Floe agent skill | npx skills add floe-labs/agent-skills |
| MCP — hosted MCP server (65 tools) | npx -y add-mcp https://mcp.floelabs.xyz/mcp |
| CLI — the full platform from your terminal: agents, keys, budgets, billing | npx @floelabs/cli init |
NPM — the SDK + floe-agent CLI |
npm i -g floe-agent |
New accounts get a $3 Welcome Credit (300 API credits) — no card. Set up with your AI tools → · Get a key →
A cookbook of small, self-contained reference agents that show Floe in real frameworks — Vapi, LangChain, CrewAI, Coinbase AgentKit, MCP, the OpenAI Agents SDK, and plain SDKs. Each folder is an independent example you can read, copy, and run on its own. This is a gallery, not a template — there's nothing to clone wholesale; pick the recipe closest to what you're building.
Floe is the spend/billing layer for AI agents. It's walletless and priced in USD. One Floe key pays every vendor your agent uses — the model, memory, tools, and 2,000+ x402 API services — through a single endpoint, governed by programmable, context-aware spend controls. That's unified billing for agents: rail-agnostic, one key, one ledger, no per-vendor accounts, no crypto.
Start free. Create an agent key at the dashboard, fund it with a card, and your agent makes its first paid API call in minutes.
- Voice agents that pay per-lookup for live web search and stop cleanly at a budget —
vapi-voice-agent,vapi-venice-voice-agent— or that remember callers across calls via HydraDB —hydra-memory-agent - Budget-capped multi-agent crews where a runaway loop dies at $1, not $414 —
crewai-demo - Metered LLM access — any OpenAI/Anthropic model behind one billed endpoint with a server-side cap —
metered-llm - Add Floe to an agent you already have — route an existing STT→LLM→TTS agent's spend through Floe's keyless gateway by swapping three values (
baseURL/apiKey/model), no provider key —drop-in-existing-agent - Zero-install access from Claude Desktop / Cursor via hosted MCP —
mcp-demo - Budget-enforce a self-hosted Pipecat / LiveKit agent — self-report each call's cost so budgets true up between sessions —
pipecat-livekit-reconcile - Graduate a Vapi / Retell / Bland agent to 100% coverage — move every leg (LLM, STT, TTS, telephony) off the orchestrator onto Floe rails so the whole call is gated pre-call, not reconciled after —
migrate-to-full-coverage
| Example | Language | Framework / Stack | Difficulty | What it shows | Link |
|---|---|---|---|---|---|
| metered-llm | TypeScript · Python | OpenAI SDK (framework-agnostic) | Beginner | Route any OpenAI/Anthropic model through Floe's metered proxy — per-token billing on one key, capped server-side, your provider key never stored. | → |
| drop-in-existing-agent | TypeScript · Python | Standard openai SDK (framework-agnostic) |
Beginner | Add Floe to an agent you already have — swap baseURL/apiKey/model to route an existing STT→LLM→TTS agent's LLM leg through Floe's keyless gateway, no provider key. |
→ |
| x402-client | TypeScript | Coinbase AgentKit | Beginner (Preview) | The minimal payment example: delegate credit to the Floe facilitator, then call any x402 API with automatic, gas-free payment. | → |
| mcp-demo | Config only | Claude Desktop / Cursor (MCP) | Beginner | Connect Claude Desktop or Cursor to Floe's hosted MCP server in one line — create agents, cap spend, and make paid calls with zero install. | → |
| openai-agents | Config only | OpenAI Agents SDK | Beginner (Preview) | Use Floe from the OpenAI Agents SDK today via MCP fallback, ahead of the native adapter. | → |
| crewai-demo | Python | CrewAI · crewai-floe |
Intermediate | Per-agent budgets with a hard, server-side ceiling: a rigged loop halts at $1, and a procurement crew enforces allowlists and per-role caps. (Installs crewai-floe from a git branch until it lands on PyPI — see its README.) |
→ |
| vapi-voice-agent | TypeScript | Vapi · GPT-4o · ElevenLabs · Exa | Intermediate | An outbound voice concierge that pays for live web search through Floe, tapers as it nears its budget, and audibly hard-stops at the cap. | → |
| vapi-venice-voice-agent | TypeScript | Vapi · Venice · ElevenLabs · Exa | Advanced | Same voice concierge, but the LLM inference itself runs on Venice through Floe — model and tools metered on one key, with an audible hard-stop. | → |
| hydra-memory-agent | TypeScript | Vapi · Venice · HydraDB | Advanced | A voice concierge with persistent memory — it stores caller facts in HydraDB and recalls them, so a later call greets you by name. Brain (Venice) + memory (HydraDB) on one Floe key, no vendor keys. | → |
| vapi-custom-llm | TypeScript | Vapi (custom LLM) | Intermediate | Point a Vapi assistant's custom LLM at Floe's keyless gateway — the model leg (~60% of a typical call's cost) is metered per token and refused 402 pre-call at the cap; an optional shim turns the refusal into a spoken goodbye. |
→ |
| retell-custom-llm | TypeScript | Retell (custom LLM WebSocket) | Intermediate | The WebSocket adapter Retell's custom LLM actually needs: a ~150-line server that turns each Retell turn into a streaming Floe gateway call — metered, pre-call gated, speaking "I've reached my budget limit" and hanging up cleanly on 402. |
→ |
| vapi-voice-metered | TypeScript | Vapi (custom transcriber + voice) · Deepgram · ElevenLabs | Advanced | Put Floe inside Vapi's provider slots: STT and TTS run through Floe's orchestrator surfaces, metered per audio-second / per character — combined with vapi-custom-llm, every leg except telephony is pre-call gated on one ledger. |
→ |
| livekit-voice-agent | Python | LiveKit Agents · Deepgram | Intermediate | A real-time LiveKit voice agent whose LLM and TTS legs meter on one Floe key via a base_url swap — STT ships BYO Deepgram, with Floe's streaming STT WebSocket as the drop-in upgrade. |
→ |
| pipecat-livekit-reconcile | Python | Pipecat · LiveKit Agents (self-hosted) | Intermediate | Reconcile Mode for self-hosted voice agents: no platform cost webhook exists, so your code POSTs each call's cost (signed) at session end — Floe enforces the budget on the next session. The better path (pre-call gating) is routing legs through Floe; this closes the gap on legs it doesn't carry. | → |
| migrate-to-full-coverage | Guide | Pipecat / LiveKit on Floe | Intermediate | The "graduate to 100% coverage" path: move each leg (LLM, STT, TTS, telephony) off Vapi/Retell/Bland onto Floe rails, flipping every reconciled leg to pre-call enforcement. Stitches the full-stack recipes with a cost/coverage calculator. | → |
| floe-phone-sales-agent | TypeScript | Floe Phone · keyless LLM · Exa · Calendly | Advanced | The full-dogfood outbound sales agent on Floe Phone (webhook mode): telephony, STT, TTS, LLM turns, and paid research all meter on one key, with enforced per-call budgets plus a campaign cap — and bookings confirmed by Calendly webhook, never on the agent's say-so. | → |
Difficulty is a rough guide: Beginner = a key and a few minutes; Intermediate = a webhook, a framework, or a running server; Advanced = a multi-vendor stack (e.g. Venice inference, HydraDB memory). (Preview) marks examples whose script is a printed walkthrough of the flow — full docs and env setup, but not yet runnable against the live API. Runnable versions are on the way; each README says exactly what runs today.
Two of the most-requested flows, end to end. Click a card to watch (~90s each).
A phone call where the LLM inference and the paid web-search tools both meter
on a single Floe key — tapering as the budget runs down to an audible hard-stop.
See vapi-venice-voice-agent.
The agent stores caller facts in HydraDB and recalls them on a later call — brain
(Venice) + memory (HydraDB) on one Floe key, no vendor keys.
See hydra-memory-agent.
Watch the setup end to end (~90s each) — key, funding, and a first metered call:
Every example is standalone. Pick one, cd into it, and follow its README.
git clone https://github.com/Floe-Labs/floe-cookbook.git
cd floe-cookbook
# Pick an example — e.g. metered LLM calls through Floe's billed proxy
cd metered-llm
cp .env.example .env # fill in FLOE_API_KEY (and any keys the README lists)
# TypeScript examples
npm install && npm start
# Python examples
pip install -r requirements.txt && python main.pyYou'll need a Floe API key — get one at the dashboard and fund it with a card. No wallet or crypto required.
Prefer to drive Floe from your coding agent? The floe agent skill
teaches Claude Code, Cursor, or any Agent Skills client to run the whole vendor stack
on one Floe key — metered, budget-capped, with the per-call receipt — the same pattern
these recipes demonstrate.
npx skills add floe-labs/agent-skillsCanonical home (issues, releases, contributions): Floe-Labs/agent-skills.
New recipes are welcome. See CONTRIBUTING.md and the per-example README template at docs/EXAMPLE_TEMPLATE.md.
MIT © Floe Labs