Guides

Using AgentRail Chat

A practical guide to AgentRail Chat — the conversational interface to your wallets, agents, and payments. Built into every account.

Where you'll find it

AgentRail Chat ships with every account. There are three surfaces:

  • ─Console — Chat lives in the right rail of every page in the AgentRail console, scoped to whichever wallet, agent, or transaction you're looking at.
  • ─Mobile — the AgentRail mobile app puts Chat front and center, so you can approve a payment from the airport or pause an agent from a Lyft.
  • ─Slack — install the AgentRail Slack app, then @AgentRail anywhere in your workspace. Mentions route to the right channel for audit.

Scope and permissions

Chat acts on your behalf using your AgentRail credentials. It can do anything you can do — no more, no less. If your account doesn't have payments.create permission, neither does Chat.

You can also scope Chat to a single agent, a single wallet, or a single environment. Many teams restrict mobile Chat to read-only + approvals, and keep mutating actions to the desktop console.

Heads up
Chat does not bypass your spend policies. Every prompt runs through the same policy envelope your agents do.

Prompt patterns that work

Chat is best at specific intents tied to AgentRail primitives. Be concrete about amounts, counterparties, and wallets — the more specific the prompt, the less Chat has to ask back.

prompt
# Specific — Chat acts immediately
Pay Acme Logistics $4,250 from the dispatch wallet, memo "Run 2026-05-19".

# Slightly less specific — Chat will ask which wallet
Pay Acme Logistics $4,250 for the May 19 run.

# Open-ended — Chat will scope down with questions
Reconcile yesterday's carrier payments.

Chat understands references to objects by name. You don't need to paste IDs — “the dispatch wallet,” “the carrier payment agent,” or “Acme” resolve as long as they're unambiguous within your scope.

Approvals inline

When a prompt triggers a policy that requires approval (over a cap, a sanctions hit, a new counterparty), Chat surfaces an approval card inline. One tap clears it — and the audit trail records who tapped, from which device, at what time.

On mobile, approval cards use Face ID / Touch ID by default. In Slack, approvals are message-attached and resolve via reaction or slash-command.

Every prompt is a receipt

Every Chat turn — the prompt, the resolved action, the policy verdict, the resulting transactions — is recorded as one or more events on the Observability stream. You can replay any chat session weeks later and reconstruct exactly what happened and why.

What Chat won't do

  • ─Chat won't bulk-import payments without explicit confirmation.
  • ─Chat won't edit your spend policies — that's console-only, on purpose.
  • ─Chat won't talk to systems outside AgentRail. It's a UI on top of our API, not an open-ended assistant.
  • ─Chat won't share data across organizations. Your scope is your scope.
← Back to docs indexAsk AgentRail Chat →