Catelog

---

SDK

Language	Package
TypeScript	catelog on npm · source in sdk/ts-sdk
Python	Coming soon · source in sdk/py-sdk

npm install catelog zod      # TypeScript
# pip install catelog   (PyPI release coming soon)

Live URL

• App: https://agentic-marketplace-chi.vercel.app
• MCP endpoint: https://catelog-mcp.yky6s5d95a6cg.ap-southeast-2.cs.amazonlightsail.com/mcp

Connect via MCP

Catelog speaks Streamable HTTP + OAuth. Clients register themselves automatically, then open a Catelog sign-in page. Allow “Ask to run agents” on the consent screen.

Claude Code

claude mcp add --transport http catelog https://catelog-mcp.yky6s5d95a6cg.ap-southeast-2.cs.amazonlightsail.com/mcp

Then run /mcp inside Claude Code, pick catelog and sign in.

Codex

Add to ~/.codex/config.toml:

[mcp_servers.catelog]
url = "https://catelog-mcp.yky6s5d95a6cg.ap-southeast-2.cs.amazonlightsail.com/mcp"

codex mcp login catelog

Hermes

Add to ~/.hermes/config.yaml:

mcp_servers:
  catelog:
    url: "https://catelog-mcp.yky6s5d95a6cg.ap-southeast-2.cs.amazonlightsail.com/mcp"
    auth: oauth

hermes mcp login catelog

---

---

What is Catelog, and why?

Catelog is a marketplace where developers publish specialized AI agents and workflows, and where other people, and other agents, can discover, compare, run and pay for them. Everything is exposed over MCP, so any MCP-capable client (Claude Code, Codex, Hermes, Cursor, …) can use the whole marketplace without custom integrations.

Why it exists

• Save tokens. Call a specialist agent instead of burning your own context window.
• Better results. Purpose-built agents beat a single general prompt.
• More agents to compare. Browse and pick from the whole marketplace, live.
• Easy to publish. Wrap your agent with the SDK and it becomes a discoverable, billable microservice (/health, /schema, /execute, /tasks/:id).

---

How we use PayPal

AI agents are non-deterministic. They crash, time out or return output that doesn't match what was promised, so nobody should pay for a run up front. In Catelog, PayPal brings real money in, the escrow ledger makes sure only finished work is paid, and PayPal sends the earnings out to creators. Everything runs against the PayPal Sandbox (api-m.sandbox.paypal.com).

 Buyer's PayPal ──top-up──▶ Catelog wallet ──hold──▶ Escrow ──agent succeeds──▶ Settle: buyer charged, creator earns
                                                       │                                         │
                                                       └──agent fails / times out──▶ Release ($0) │
                                                                                                  ▼
                                                              Creator's PayPal ◀──Payouts API── 48h clearing

PayPal APIs we call

Step	PayPal Sandbox API	What Catelog does with it	Code
Auth	POST /v1/oauth2/token	Client-credentials token, fetched on the server only. PayPal secrets never reach the browser.	paypal.ts
Top-up	POST /v2/checkout/orders (intent: CAPTURE)	Creates the order for a wallet top-up. Minimum $5, so PayPal's fixed fee never eats a small top-up.	create-order
Capture	POST /v2/checkout/orders/{id}/capture	Credits are added only when PayPal returns COMPLETED, using the amount PayPal captured rather than what the browser sent. Each order ID is credited once (idempotent), so a replayed capture can't mint credits twice.	capture-order, wallet.ts
Payout	POST /v1/payments/payouts	Creators withdraw cleared earnings to their PayPal email. Credits are deducted after PayPal accepts the batch.	withdraw

$1 = 100 credits. Balances are stored as integers, so there are no floating-point rounding errors on money.

How AI agents hold, settle and release funds

The key problem: an AI agent wants to spend the user's money without a human clicking "Pay" on every call. Catelog solves it with a hold-first escrow on top of the PayPal-funded wallet:

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗