MCP Overview

Panoptes serves a Model Context Protocol server so an AI agent — Claude, ChatGPT, Cursor or your own — can answer questions about a restaurant from real numbers and, when you allow it, take actions.

URL	https://api.panoptesos.com/api/v1/mcp (staging: https://staging-api.panoptesos.com/api/v1/mcp)
Transport	Streamable HTTP, stateless JSON-RPC
Auth	X-Api-Key header, or an OAuth bearer from the Panoptes OAuth issuer
Admission	cap:view_reports (or the legacy read:reports)

In the dashboard, Settings → Developers → MCP / AI shows the URL for your install, ready-to-paste client configs and the tool list.

How an agent works with it

1. get_business_context first. It names the branch the credential is bound to, the currency every amount is in, today's business date and the grounding rules.
2. Facts answer the common questions (sales, prime cost, COGS, loss, shrinkage, reorder, payroll…) with numbers computed by the server — the agent never does the arithmetic.
3. Tables let it look further: list, read, count and aggregate the readable tables.
4. Actions let it change data: list_actions finds what the credential may call, call_action performs it through the REST pipeline.

Guarantees

• Same permissions as the REST API. Effective access is the owner's role AND the credential's scopes. Each fact tool keeps its own capability.
• Reads cannot write. Fact and table tools run in a READ ONLY transaction; call_action is the only write path.
• Writes are the real thing. call_action runs in-process through the console pipeline: validation, separation of duties, period locks, idempotency and the audit chain all apply.
• The model is told to ask. The server instructs the agent to confirm any money or stock movement with the user, and to pass an idempotencyKey on every write.

Next: Connect a client · Tool reference · Actions