Live edge API | remote MCP | npm package

One OTP.
Zero API keys.
Agents that can actually do the job.

AgentPay is the trust and payment layer for AI agents. Vault credentials with one hosted connect flow, enforce governed mandates, and settle paid actions through one MCP server for Claude, OpenAI, and any MCP-compatible host.

Read quickstart Register in browser View GitHub

npx -y @agentpayxyz/mcp-server

Remote MCP endpoint: https://api.agentpay.so/api/mcp

First install

30 seconds from config to first governed tool call

Drop this into Claude Desktop. Restart. Ask the host to connect Firecrawl or create a mandate. AgentPay handles credential vaulting, budget enforcement, payment recovery, and receipts behind the scenes.

Connect Firecrawl without exposing the raw API key to the agent.

Create a governed mandate to scrape this site with a $5 budget cap.

Create a funding request so the human can pay without leaving the chat.

~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "agentpay": {
      "command": "npx",
      "args": ["-y", "@agentpayxyz/mcp-server"],
      "env": {
        "AGENTPAY_API_KEY": "apk_your_key"
      }
    }
  }
}

Three hard problems. One product surface.

This is where the value lives today. Keep the public story tight: protect credentials, govern spending, and return proof that the action settled.

Capability Vault

Zero API keys in agent context

Users approve one hosted connect flow. AgentPay vaults Firecrawl, Perplexity, OpenAI, and other upstream credentials so the raw key never enters the prompt loop.

agentpay_request_capability_connect

Governed Mandates

Agents that spend within limits

Your agent proposes the action, budget cap, and approval threshold. The human approves once. AgentPay enforces the policy on every subsequent step.

agentpay_create_mandate

Settlement

Payment recovery and receipts built in

Collect card or UPI funding, execute the action, and return a verifiable receipt. The host sees a clean workflow instead of bespoke payment plumbing.

agentpay_create_human_funding_request

How it works

The activation loop

AgentPay should feel like the safe execution layer behind the host, not another dashboard the user has to babysit. The install path, connect path, approval path, and receipt path need to be obvious on first use.

Install the MCP server

Add AgentPay to Claude Desktop or any MCP host with one command and one API key.

Connect paid capabilities once

When the agent needs Firecrawl, Perplexity, or another upstream service, AgentPay opens the hosted connect flow and vaults the credential.

Approve the mandate

The agent proposes the action and budget. The human approves once. Policy is locked before execution starts.

Let the agent finish

AgentPay enforces the budget, runs the action, and returns the settlement and receipt trail the developer needs.

Works with any MCP-compatible host

Claude Desktop

Claude Code

GPT-4o

Cursor

Any MCP host

Or connect via remote MCP at https://api.agentpay.so/api/mcp

What teams can ship this week

The product face should stay on the wedge.

No ACE or RCM dependency here. The public story should stay on the installable infrastructure developers can evaluate quickly: governed browsing, paid execution, and safe upstream capability access.

Paid research agent

Give Claude or GPT a governed path to browse, scrape, and summarize without storing the user's Firecrawl key in the prompt loop.

Zero-key browsing copilot

Ship Browserbase or Firecrawl access through AgentPay instead of telling every user to manage their own upstream credentials and billing.

Chat-native checkout

Create a funding request in chat, collect payment, and let the agent continue the workflow without tab switching or manual settlement glue.

api.agentpay.so/api/mcp