PayPilot — agentic commerce operator for PayPal

Natural language in, PayPal operations out. PayPilot is a solo-built entry for the PayPal AI Hackathon, targeting the main prizes and Best Use of Agentic Commerce.

Say "Invoice Acme Corp $1,250 due in 30 days", "Who hasn't paid yet?", or "Refund the last payment from John" — a tool-calling agent turns it into real PayPal REST API calls through a self-hosted MCP server.

What it does

• 8 PayPal tools over a self-hosted MCP server (Streamable HTTP): invoice_create, invoice_send, invoice_remind, invoice_list, invoice_status, payment_search, refund_payment, account_summary

• Agent loop (paypilot/agent.py): a compact ReAct-style tool-calling loop over any OpenAI-compatible chat endpoint. No heavy framework, no silent fallbacks — an unconfigured model raises a clear error.

• PayPal REST client (paypilot/paypal_client.py): OAuth2 client-credentials, Invoices v2 (create/get/list/send/remind/cancel), Transaction Search (payment lookup), Payments v2 refunds. All network I/O goes through an injectable transport.

• Mock transport (paypilot/transport.py): an in-memory PayPal simulation (OAuth, invoice lifecycle with real state transitions, transaction search, refunds) so the whole suite runs without credentials.

• Web demo (paypilot/web.py + webui/index.html): chat UI that shows every tool call the agent makes.

Related MCP server: PayPal MCP Server

Quickstart

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# run the tests (mock backend, no credentials needed)
python -m pytest

# start the MCP server (mock backend by default)
PAYPILOT_TRANSPORT=mock python -m paypilot.mcp_server

# start the web demo (needs a model; see below)
PAYPILOT_TRANSPORT=mock \
PAYPILOT_MODEL_BASE_URL=http://127.0.0.1:11435/v1 \
PAYPILOT_MODEL_ID=nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B \
python -m paypilot.web
# open http://127.0.0.1:8931/

Real PayPal sandbox setup

1. Go to https://developer.paypal.com/dashboard/applications/sandbox and log in with a PayPal developer account.

2. Create an app (or open the default one) and copy the Client ID and Secret.

3. Export them:

   export PAYPAL_CLIENT_ID="your-sandbox-client-id"
   export PAYPAL_CLIENT_SECRET="your-sandbox-client-secret"
   export PAYPAL_MODE=sandbox   # the default; use "live" only deliberately

4. Run the live smoke test (full invoice lifecycle against api-m.sandbox.paypal.com, with cleanup):

   python scripts/sandbox-smoke.py

   Without credentials it exits 2 with these instructions; nothing is fabricated. To also smoke-test refunds, set PAYPAL_SMOKE_CAPTURE_ID to a real sandbox capture id (refunds need a completed payment).

Unset PAYPILOT_TRANSPORT (or set it to anything but mock) to use the real sandbox backend in the MCP server and web demo.

Model configuration

The agent needs an OpenAI-compatible chat-completions endpoint with tool calling:

export PAYPILOT_MODEL_BASE_URL="http://127.0.0.1:11434/v1"
export PAYPILOT_MODEL_API_KEY="..."   # if the endpoint needs one
export PAYPILOT_MODEL_ID="model-id"

Without it, the agent raises ConfigurationError naming the missing variables instead of pretending to work.

Project layout

paypilot/
  paypal_client.py   PayPal REST client (OAuth, invoices, search, refunds)
  transport.py       UrllibTransport (real) + MockPayPalTransport (tests)
  tools.py           shared tool specs + dispatcher (agent/MCP/web)
  agent.py           tool-calling agent loop + model clients
  mcp_server.py      self-hosted MCP server (Streamable HTTP, 8 tools)
  web.py             FastAPI demo UI backend
  netenv.py          localhost proxy-env sanitizer (test env fix)
webui/index.html     demo chat page
tests/               full suite (mock backend; MCP over real HTTP)
scripts/sandbox-smoke.py  live sandbox lifecycle test

License

MIT. Built solo for the PayPal AI Hackathon.