AgentPay

AgentPay is the authority layer between an AI agent and the real world. It handles provider access, funding authority, human approval, exact-call resume, and proof so agents can finish work end to end without raw secrets in chat or repeated dashboard setup.

It is not a generic "agent platform." It is the trust seam that makes autonomous execution usable.

npx -y @agentpayxyz/mcp-server

That one command gives any MCP-compatible host the ability to:

• request API access

• vault credentials without exposing the raw key to the agent

• run governed paid execution

• pause for OTP or approval only when needed

• resume the exact blocked call automatically

• reuse governed access later from the same workbench

AgentPay Commerce Network

AgentPay now has a focused ecommerce surface: shoppers discover products by a concrete need, then set a visual Buyer Constitution for category scope, budget, delivery, returns, merchant policy, and human approval. AgentPay enforces that constitution before optional model ranking, and approved matches continue to merchant checkout. Merchants get the other side of the demand graph through catalog truth, source-grounded product imagery, and signed attribution drafts.

• Shopper experience: /commerce

• Seller growth workspace: /commerce/studio

• Need-led API: POST /api/commerce/discover

• Closed-world GPT-5.6 compiler: POST /api/commerce/compile

• Catalog audit API: POST /api/commerce/catalog/audit

• MCP tools: agentpay_discover_products, agentpay_compile_product_decision, and agentpay_audit_catalog_truth

• Founder and SEO specification: docs/COMMERCE_CONTROL_PLANE.md

• OpenAI Build Week evidence and judge runbook: docs/OPENAI_BUILD_WEEK_2026.md

• Dated hackathon reuse decisions and event-specific build boundaries: docs/HACKATHON_REUSE_2026-07-18.md

The included inventory and demand indexes are sandbox fixtures. They do not represent live merchants, customers, orders, or revenue. Merchant connections, public inventory, conversion webhooks, fee collection, and deployment remain separate approval and validation gates.

Related MCP server: agent-sudo-mcp

Casper Buildathon Entry

Bee + Clickey: Founder-in-a-Box with Casper-Verifiable Agent Payments

Bee is the operator brain. Clickey is the desktop companion. Together they route work across a real local AI fleet, discover tools through the AgentPay Feed MCP, and stage payment receipts through founder-approved rails. For the Casper Agentic Buildathon, the project adds a Casper-native receipt path: Bee can stage x402-casper payment payloads and bind a Bee mandate/receipt id to a Casper Testnet transfer.

Start here:

• Submission packet

• Post-deadline recovery packet

• Bee control plane

• Casper receipt runner

Judges can verify the local proof path:

node --test ops/mac-mini/bee/bee.test.mjs
ops/mac-mini/bin/bee doctor
ops/mac-mini/bin/bee caps
ops/mac-mini/bin/bee feed
npm --prefix ops/mac-mini/bee/casper run proof
npm --prefix ops/mac-mini/bee/casper run video

Broadcasting a real testnet proof requires a dedicated funded Casper testnet key:

CASPER_PRIVATE_KEY_HEX=<testnet-private-key-hex> \
CASPER_RECIPIENT_PUBLIC_KEY_HEX=<recipient-public-key-hex> \
BEE_CASPER_RECEIPT_ID=<bee-mandate-id-or-demo-receipt-id> \
npm --prefix ops/mac-mini/bee/casper run receipt

Bee can prepare and stage high-impact actions, but it does not autonomously publish, upload, change OAuth/account settings, move money, or bypass founder approval. Live settlement remains founder-triggered.

Why it exists

Agents fail at the same place every time:

1. they need an API the developer has not wired yet

2. they need a credential the human should not paste into chat

3. they hit a paid step and lose continuity

AgentPay fixes that seam.

Capability Vault Users connect a provider once through AgentPay. The raw credential is vaulted server-side. The agent receives only governed capability access.

Governed paid execution Humans set guardrails once: funding rail, auto-approve limit, OTP policy, and spend limits. AgentPay enforces those rules when the agent acts.

Exact-call resume When a paid step needs a human, AgentPay pauses, collects the minimum approval, and resumes the exact blocked call without asking the agent to reconstruct it.

Same-workbench reuse Local projects never need raw provider keys. AgentPay can issue opaque, revocable workbench leases so the same workbench can reuse governed access later.

The inevitable path

This is the path AgentPay is built around:

1. Agent asks for an API.

2. AgentPay checks whether governed access already exists.

3. If not, AgentPay runs one hosted setup flow for authority and provider connection.

4. The agent uses the capability for free until paid usage is required.

5. AgentPay pauses for OTP or confirmation only if policy requires it.

6. AgentPay resumes the exact blocked call.

7. The same workbench can reuse governed access later without re-entering the secret.

Get started in 30 seconds

Add AgentPay to Claude Desktop (claude_desktop_config.json):

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

Get your API key:

curl -s -X POST https://api.agentpay.so/api/merchants/register \
  -H "Content-Type: application/json" \
  -d '{ "name": "My Agent", "email": "you@example.com" }'

Now ask your host:

"My agent needs Firecrawl. Set a $5 auto-approve limit, ask for OTP above that, and keep the key out of chat."

Or:

"My agent needs Databento for this workbench. If access already exists, reuse it. If not, start the minimal AgentPay setup flow."

Full quickstart

Terminal-native control plane

AgentPay should be operated through hosts and terminals, not a merchant dashboard.

Key surfaces:

• GET /api/capabilities/authority-bootstrap

• POST /api/capabilities/authority-bootstrap

• POST /api/capabilities/access-resolve

• POST /api/capabilities/onboarding-sessions

• POST /api/capabilities/lease-execute

• GET /api/capabilities/leases

• POST /api/capabilities/leases/:leaseId/revoke

• POST /api/capabilities/:capabilityId/execute

These let a host or agent:

• read authority state

• set guardrails

• connect providers

• request human approval only when needed

• reuse governed access safely later

Flagship provider paths

The current wedge is strongest when AgentPay owns setup and continuity for high-value agent APIs.

Current priority paths:

• Databento

• Firecrawl

• Browserbase

• Exa

• Generic REST API fallback

The product goal is simple: visiting provider dashboards should become the exception, not the default.

Remote MCP

For hosts that support remote MCP:

https://api.agentpay.so/api/mcp

Authenticate with your API key as a Bearer token, or mint a short-lived token for Claude, OpenAI, or another remote MCP host:

curl -X POST https://api.agentpay.so/api/mcp/tokens \
  -H "Authorization: Bearer apk_your_key_here" \
  -d '{ "audience": "openai", "ttlSeconds": 3600 }'

Ace - built on AgentPay

Ace is a proof front door built on AgentPay. It demonstrates the core seam under real-world conditions. It is not the core story.

Product truth

If a human still has to:

• paste a raw provider key into chat

• rebuild a blocked call after payment

• keep reopening provider dashboards

• or lose continuity between approval and execution

then the product is still unfinished.

Developer resources

Repository layout

apps/
  api-edge/         Cloudflare Workers public API
  meridian/         Ace front door

packages/
  mcp-server/       Published MCP package
  sdk/              TypeScript SDK
  sdk-node/         Node.js SDK

examples/           Example agents and adapters
docs/               Architecture and product notes
migrations/         PostgreSQL migrations
ops/                Founder and growth operating artifacts

License

Business Source License 1.1. Converts to AGPL-3.0 on 2029-01-01.