floe-agent

· ·

The spend layer for AI agents — TypeScript SDK. Pay any of 2,000+ vendor API services through one endpoint, with budgets your agent can reason about. Walletless. No crypto required.

Website · Docs · Dashboard · Python SDK

What it does

Your agent calls paid APIs — LLMs, voice, search, data. Floe is the spend layer in front of those calls:

• One endpoint, many vendors. Pay any x402 API through the Floe facilitator. No per-vendor accounts or keys.
• Budgets the agent can reason about. Ask "do I have budget? is this call worth it?" before paying.
• Programmable spend controls. Per-call caps, daily limits, allowed destinations, session ceilings — enforced server-side, before money moves.
• Walletless. Email + a funding source. Floe provisions wallets in the background — no MetaMask, no seed phrase, no gas. The stablecoin rails are invisible.
• Real-time visibility. Every call is a typed receipt: target, amount, status, time.

Two layers in one package:

Install

FloeAgent (the runtime client) can be used standalone:

npm install floe-agent

The AgentKit action providers need @coinbase/agentkit and its peers:

npm install floe-agent @coinbase/agentkit viem zod

30-second example

No wallet, no chain. Your agent pays for an API and stays inside its budget.

import { FloeAgent } from "floe-agent";

const agent = new FloeAgent({ apiKey: process.env.FLOE_API_KEY });

// Check budget before spending. `balance()` returns a plain number.
const available = await agent.balance();   // dollars available to spend
// For active loans, pending settlements, and raw units, use:
//   const detail = await agent.balanceDetails(); // BalanceResult

// Pay any x402 API. The agent never touches the payment layer.
const res = await agent.fetch({
  url: "https://api.exa.ai/search",
  method: "POST",
  body: JSON.stringify({ query: "AI agent frameworks" }),
});
// res is a FetchResult: res.body (string), res.cost (USDC dollars), res.status

Fund with fiat: operators fund the wallet with USDC via Coinbase — card, bank transfer, Apple Pay, Google Pay — directly from the dashboard. No crypto on-ramp.

Get an agent key: dashboard → connect → Create an agent → copy the floe_<hex> key (shown once).

The decision loop (what makes this different)

Most SDKs let an agent spend. Floe lets an agent reason about spend before it commits, then stops it before it overruns.

// 1. Preflight the cost against your balance — no payment is made.
const estimate = await agent.estimateCost("https://api.example.com/premium");
// estimate => { cost: number; canAfford: boolean; isPaid: boolean }

// 2. Only pay if the agent can afford it.
if (estimate.canAfford) {
  await agent.fetch({
    url: "https://api.example.com/premium",
    method: "POST",
    body: JSON.stringify({ prompt: "..." }),
  });
}

This is the "answer the three rational-agent questions in one round-trip" workflow: Do I have budget? Is this call worth it? Where am I in my spend?

The lower-level x402ActionProvider action estimate_x402_cost is a separate API that returns a richer reflection (willExceedAvailable / willExceedSpendLimit) for AgentKit tool calls — see the Advanced section.

Framework support

// Vercel AI SDK (requires: npm install floe-agent @coinbase/agentkit viem zod)
import { AgentKit } from "@coinbase/agentkit";
import { getVercelAITools } from "@coinbase/agentkit-vercel-ai-sdk";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
import { floeActionProvider, x402ActionProvider } from "floe-agent";

const agentkit = await AgentKit.from({
  walletProvider,
  actionProviders: [
    floeActionProvider(),
    x402ActionProvider({ facilitatorApiKey: process.env.FLOE_FACILITATOR_API_KEY }),
  ],
});

const tools = await getVercelAITools(agentkit);
const { text } = await generateText({
  model: openai("gpt-4o"),
  tools,
  maxSteps: 10,
  prompt: "What did my agent spend today?",
});

// MCP (Claude Desktop / Claude Code / Cursor) — zero install
{ "mcpServers": { "floe": {
  "url": "https://mcp.floelabs.xyz/mcp",
  "transport": "streamable-http"
} } }

npm install floe-agent @coinbase/agentkit viem zod

import { AgentKit } from "@coinbase/agentkit";
import { floeActionProvider, x402ActionProvider } from "floe-agent";

const agentkit = await AgentKit.from({
  walletProvider,
  actionProviders: [
    floeActionProvider({ knownMarketIds: ["0x..."] }),
    x402ActionProvider({ facilitatorApiKey: process.env.FLOE_FACILITATOR_API_KEY }),
  ],
});

// Borrow against on-chain collateral. marketId is required — fetch via get_markets.
await agentkit.run("instant_borrow", {
  marketId: "0x...",
  borrowAmount: "9500000000",
  collateralAmount: "10000000000",
  maxInterestRateBps: "800",
  duration: "1209600",
});

import { AgentKit } from "@coinbase/agentkit";
import { floeActionProvider, x402ActionProvider } from "floe-agent";

const agentkit = await AgentKit.from({
  walletProvider,
  actionProviders: [
    floeActionProvider(),
    x402ActionProvider({ facilitatorApiKey: process.env.FLOE_FACILITATOR_API_KEY }),
  ],
});

CLI

npm install -g floe-agent
floe-agent

Interactive conversational agent for testing the full action surface. Prompts for wallet provider (Private Key or CDP MPC), AI provider (OpenAI / Anthropic / Ollama), and Base RPC URL. Config saved to .floe-agent.json; API keys are never cached.

Environment variables (CLI / examples)

Examples

See floe-cookbook for end-to-end multi-framework agents. The examples/ directory here has a Vercel AI SDK chatbot (chatbot.ts) and a no-framework standalone script (standalone.ts).

# npm link (live symlink)
npm run build && npm link        # then `npm link floe-agent` in your project

# npm pack (simulates a real publish — safest validation)
npm run build && npm pack        # creates floe-agent-<version>.tgz
npm pack --dry-run               # should include only dist/, README.md, package.json

Roadmap

Floe ships the spend layer first and builds credit on top of it.

• Working capital / credit lines — in development.
• Unsecured / receivables-backed credit — in development. Email hello@floelabs.xyz for the design-partner program.
• Portable ERC-8004 credit record — in development.

Links

Website · Docs · Python SDK · MCP server · Examples

License

MIT