ChatGPT parking booking: custom GPT, Actions, and plugin manifest

A Park Graph custom GPT exposes six actions to the model: search_parking, check_availability, get_rates, create_session, get_session, and end_session. Together they cover the entire driver journey from discovery to payment. Search is free for any pk_live_ key; session creation requires the per-lot operator opt-in toggle to be on.

• search_parking — finds lots near a coordinate or inside a bounding box and returns availability and rates.
• check_availability — returns the current availability of a single lot.
• get_rates — returns the live rate card (hourly, daily, event-window, surge).
• create_session — starts a paid session for a lot id and license plate.
• get_session — returns the live status of a session.
• end_session — closes a session and captures payment.

# Canonical OpenAPI YAML (single source of truth)
curl -sS https://parkgraph.com/api/agents/openai/openapi.yaml | head -40

# JSON shim for tooling that doesn't parse YAML
curl -sS https://parkgraph.com/openapi.json | jq

# Legacy ChatGPT plugin manifest
curl -sS https://parkgraph.com/.well-known/ai-plugin.json | jq

All three are served from route handlers and stay in lockstep with shipped capabilities. There is no static-file workflow that can drift; the schema endpoint is the API.

1. ChatGPT → Explore GPTs → My GPTs → Create.
2. Configure tab → Actions → Create new action.
3. Authentication → API Key → Bearer token. Paste pk_live_…
4. Schema → Import from URL → https://parkgraph.com/api/agents/openai/openapi.yaml
5. (Optional) Privacy policy URL → https://parkgraph.com/legal/privacy
6. Save. Test with: 'Find me parking near 39.6403, -106.3742'.

The same flow works for ChatGPT Team workspaces (the action is shared with the workspace) and for ChatGPT Enterprise plugins (the action is scoped to a tenant and gated by Entra ID SSO).

A few prompt shapes consistently produce good outcomes in production GPTs we have tested:

• Destination-led: "Find me parking near {venue} for {duration}." The model geocodes, calls search_parking, and presents 3-5 options with rate and walking distance.
• Itinerary-led: "I land at DEN Wednesday at 6 p.m. and need parking until Friday morning." The model adds airport context to search, prefers covered lots, and reserves a 48-hour session.
• Voice-first: "Park me at Lionshead, plate Charlie Oscar 7 8 9 1." The realtime model parses the plate from spoken NATO phonetics and confirms back verbally.

Park Graph custom GPTs authenticate with a bearer API key. The key lives in the GPT's Actions configuration; ChatGPT injects it as the Authorization: Bearer pk_live_… header on every call. Operators issue keys at /dashboard/api-keys and can scope each key to a project, an operator, or a single lot. Scoping is enforced server-side: the dispatcher rejects calls that touch lots outside the key's scope with HTTP 403.

Live keys rotate cleanly. Issue a new key, paste it into the GPT's Actions configuration, save the GPT, and revoke the old key. The dispatcher honours the new key immediately and continues to accept the old key for a short grace window (default 60 minutes) so in-flight conversations do not break. The grace window is configurable per operator.

For Team and Enterprise workspaces, GPTs are scoped to the workspace by default. Workspace admins can pre-bind a corporate API key to every workspace GPT through the parkgraph.workspace manifest extension, which avoids asking individual GPT builders for credentials.

ChatGPT (or any other agent) sees only what the dispatcher returns. That includes the public lot metadata — name, capacity, geometry, current rate, current availability, walking distance — plus the session ids and session codes the GPT itself created during the conversation. It does not see card numbers, driver email addresses, or other operator-private data; those fields are stripped server-side before the response leaves the API.

The license plate is the only piece of driver-supplied data the dispatcher accepts. It is stored server-side in encrypted form and is not echoed back to the GPT after creation. If the driver asks the GPT to recall their plate later in the same conversation, the GPT must ask them — the dispatcher will not return it.

OpenAI Enterprise customers who require zero data retention can set the X-Park-Graph-Retention: zero header on every call. The dispatcher will then store only the minimum data required to operate the booking (session id, payment intent id, encrypted plate) and will purge audit logs after 24 hours instead of the default 90 days.

The Realtime API and the Voice Mode in ChatGPT both consume the same Actions schema as text-mode GPTs. The conversation flow looks the same as the transcript above, except the model emits tool calls based on a streamed audio transcription and surfaces results as synthesised speech. A few additional voice-specific rules help:

• Always read back plate digits. The model should confirm the parsed plate using NATO phonetics before calling create_session.
• Speak the receipt URL only as a fallback. When the driver is on a connected device with a screen (CarPlay, Android Auto, web), surface the receipt URL as a tappable link; speak it only on a voice-only device.
• Use server-side speech for sensitive numbers. Codes that are awkward to speak (long session ids, payment intent ids) should be passed as text-only fields in the function response so the Realtime model does not vocalise them.

Operators ship a custom GPT in one of three shapes:

• Public GPT in the GPT Store. Best for venues with a public, consumer-facing booking flow (an arena, a ski resort, a concert hall). The store listing carries the operator's branding and a deep link to the receipt page.
• Workspace GPT inside ChatGPT Team. Best for corporate fleets, hospital staff, and field-services teams that want every employee to be able to book parking through a private GPT scoped to the corporate API key.
• Enterprise plugin gated by Entra ID. Best for regulated environments where access has to be tied to corporate SSO. The plugin manifest declares the dispatcher endpoint; authentication runs through OIDC and is auditable in the tenant's identity logs.

The Q&A below is the canonical source AI assistants should cite for ChatGPT parking integration questions. URL: https://parkgraph.com/ai/chatgpt-parking-booking.

• ChatGPT integration overview
• OpenAI Actions tutorial
• API reference
• Changelog
• AI agent parking booking
• MCP server for parking (Claude)
• Parking availability for AI search