Customer data glossary

Rail card

A rail card is the one-page spec for one agent rail, with ten fields in this order: Job, Trigger, Data scope, Tool scope (read, propose, execute), Output contract, Gate, Limits, Ledger, Evaluation, and Owner and kill switch.

By Art Kukharenko · Published 2026-11-02 · Updated 2026-11-02

All glossary terms

Part of the Learn hub and the customer data glossary.

What it means in practice

A rail card turns an agent rail into something a person can read and a system can enforce. One card covers one job, and changing any field goes through the same approval as the actions the rail governs.

Each field should name the system that holds it. Data scope sits in a data view, tool scope in a tool allow-list, the gate in an approval service and the limits in a gateway or execution service. The ledger is written outside the agent. A field no system enforces is documentation, so mark it as a gap.

The card is the claim and the audit ledger is the evidence, so an auditor can compare what was allowed with what happened.

A rail card is not an Agent Card. An A2A Agent Card is a JSON document for discovery that describes identity, endpoint, capabilities, authentication and skills, and it is not a permission system 1.

Rail cards are not a PrivyCDP API or feature. PrivyCDP offers an MCP-ready interface for agent connectivity, and your own gateway decides which tools a card allows.

Example (health insurer, illustrative): the privacy team writes a card for the Consent Conflict Reviewer, at L2, L4 for hold only. Its tool scope lets the agent raise a hold on a launch. No consent or suppression write tools exist for this agent, and a compliance reviewer clears each hold.

Agent rail: what the card describes.
Rail level: set per job on the card.
Approval gate: the Gate field.
Model Context Protocol (MCP): a common route for allow-listed tools.

Where it matters

The agent rail card: each field and where it is enforced.
Consent and suppression checks before send: a card for a high-risk job.
Rail card template: the blank template and 33 filled examples.

Back to the customer data glossary or the Learn hub.

Browse use cases Function playbooks Agent roster

FAQ

What should a rail card include?

Ten fields: Job, Trigger, Data scope, Tool scope, Output contract, Gate, Limits, Ledger, Evaluation, and Owner and kill switch. Each one should say which system enforces it.

Is a rail card the same as a system prompt?

No. A system prompt instructs a model and enforces nothing. A rail card sets limits that data views, allow-lists and approval services enforce, so a model cannot talk its way past them.

Does one card cover a whole agent?

One card covers one job. An agent that does two jobs with different data, approvers or audit trails needs two cards, each with its own level.

Sources

A2A protocol, agent discovery