MCP for AI agents

Connect an AI agent over MCP to inspect access state and trigger checks — read-only, repairs excluded.

Connect an AI agent — like Claude Code — to ParityRail over MCP (Model Context Protocol). It can then ask about a customer’s access and trigger a check, right from your chat — all read-only.

Endpoint

ParityRail’s MCP server lives at this URL. It speaks streamable HTTP, the standard MCP transport — most MCP clients handle that detail for you.

http(s)://<host>/api/mcp

Authenticate the same way you would for the REST API: send a ParityRail API key as a bearer token, using the same keys you create in Settings → API keys.

Authorization: Bearer sk_prail_...

The key controls what the agent can see, too — a project-scoped key only sees that project, and an org-scoped key sees any project in your organization.

MCP is read and scan only — an agent can inspect state and trigger a fulfillment check, but it can never approve or execute a repair, whatever the key’s scope. Repairs always go through a person in the dashboard or a repair-scoped REST call.

Connect Claude Code

claude mcp add --transport http parityrail https://your-host/api/mcp --header "Authorization: Bearer sk_prail_..."

Replace your-host with your ParityRail host and the token with your API key. Any MCP-capable client connects the same way — point it at the endpoint with the bearer header.

Tools

These tools ask and check — none of them can change anything. For example, ask your agent: “Is cus_123getting the access they paid for?” and it will check and answer from the same data your dashboard shows.

Tool	What it does
`list_projects`	List the projects the API key can access.
`get_customer_state`	Return the billing promise, live app state, and open incidents for a customer.
`list_incidents`	List open access incidents for a project.
`get_access_ledger`	Return a customer's Access Ledger — billing events, promises, checks, incidents, repairs, and verifications.
`run_check`	Trigger a read-only fulfillment check for a project or a single customer.

No tool writes to your database or to Stripe. Repairs are deliberately left out of MCP — approving and executing one is a deliberate, audited action that always happens in the dashboard or the REST API, never in chat.