Argorant Help
EN|DEargorant.comOpen Argorant
Help/AI agents & API

Spending guardrails: instructions that keep an agent cheap

Published by Argorant

Copy-ready instructions that keep an AI agent from spending credits, sending email or buying without you, plus the limits Argorant enforces on its own.

An AI agent connected to Argorant can do real things: spend credits on contacts, start exports, launch campaigns that send email and answer replies. This article gives you instructions to paste into your agent so it stops and asks at the right moments, and explains what Argorant already enforces without any instructions.

What Argorant enforces on its own

These hold no matter what your agent is told.

  • An agent can never spend more than the credits in your account. When credits run out, paid actions stop with a clear message.
  • Speed limits and daily limits apply to every connection. During the trial, each connection can reveal up to 100 people a day, and searches through an AI assistant are capped per day for the whole account.
  • Bulk email finding never spends more than the limit the agent passes for that run. It pauses at the limit and waits.
  • Checking your own list never costs more than the quote or the maximum the agent passed, whichever is smaller.
  • Ordering mailboxes needs an explicit confirmation flag from the agent. In ChatGPT and Claude, and in Codex, an order always returns a payment link that you open and pay yourself.
  • A campaign launch is refused while an email still has placeholder text like [Name] or TBD, or has no mailbox or no people.
  • Every spend is recorded and visible to you under Settings, Plans and usage.
The What costs credits section under Settings, Plans and usage, listing the price of each paid action
Settings, Plans and usage lists exactly what costs credits. Give your agent the same list.

Require explicit approval of the scope, row limit and credit budget before reveal, person enrichment or a new export, and before launching a campaign or sending a reply. Client tool labels are not a universal confirmation guarantee; non-interactive CLI commands can spend immediately.

The base rules to paste

Put this in your agent's instructions, project rules or system prompt. It works for Claude, ChatGPT, Codex, Cursor and agents you build yourself.

Rules for using Argorant:
1. Counts, previews and reading my account, lists, campaigns and replies are free. Use them freely.
2. Before any action that spends credits (reveal, export, adding people
   to a campaign, finding or enriching emails, checking an email list),
   tell me the number of people or rows, the most it can cost in credits,
   and what I get. Then wait for my explicit yes.
3. A yes covers that one action at that size. Ask again for anything bigger
   or different.
4. Never launch a campaign, send a reply or forward an email without my
   clear yes on that exact campaign or text.
5. Never order mailboxes or domains, or change my plan.
   If credits run out, tell me and stop.
6. Never retry a failed paid action or start a second export. Check the
   status first.
7. Keep contact details in Argorant. Give me links to Lists and Exports
   instead of pasting rows into the chat.

Stricter variants

Pick the one that fits the job and add it below the base rules.

A research agent that never spends

You may only count, preview and read. Never call a tool that reveals,
exports, finds emails, enriches people, adds people to a campaign or
checks email lists. If I ask for contact details, explain what it
would cost and tell me to do it in the Argorant app.

A fixed budget per session

Your Argorant budget for this session is 200 credits in total.
Keep a running total of what you spent, using the cost each result
reports. Before every paid action, show the running total and the new
maximum. Stop and ask when the next action would go over 200.

A campaign helper that never sends

You may create and edit draft campaigns, write emails and pick
mailboxes. Do not add people from Argorant's search to a campaign
without asking, because each new person costs 1 credit. Never launch,
resume or delete a campaign, and never send or forward replies.
When a draft is ready, list what still blocks a launch and stop.

An agent that uses the command line

Use npx argorant with --json. Before reveal, export, find or verify
--file, run the matching count first and show me the row limit (-n) and
the most it can cost. Wait for my yes. Never pass --yes on your own.
Branch on exit codes: 2 means sign in again, 3 means not allowed,
4 means wait, 5 means credits or a plan are needed. On 5, stop and tell me.

Settings that back up your instructions

  1. Know the balance before unattended runs. An agent cannot spend what is not there.
  2. Use your app's own approval. Claude, ChatGPT and Cursor can ask before a tool runs. Leave that on for Argorant, at least for tools that are not read-only. Argorant marks its read-only tools so they do not trigger a question each time.
  3. Disconnect what you do not use. Under Settings, AI connections, Connected apps, every app has a Disconnect button. Access ends at once.
  4. One key per job. For command line or REST agents, create a separate key under Settings, API keys, so you can delete just that one when the job is done.
Settings, AI connections with the Connected apps section, where each AI app can be disconnected
Connected apps is your off switch. Disconnect ends an app's access at once.

The command line trap

The Argorant command line asks for confirmation only when a person types at a terminal. Inside an agent, a pipe or a CI job, and with --json or --yes, there is no question. reveal, export, find and verify --file spend right away. Write your rules as if the confirmation does not exist, because for your agent it does not.

Common mistakes

  • Approving "go ahead" without a number. Ask the agent to always name rows and credits, so your yes is about something specific.
  • Saying "save them" and meaning five people. A saved list holds the whole search. Saving is free, but adding that list to a campaign costs 1 credit per new person.
  • Letting an agent loop the single email finder. Tell it to use the list version with a spending limit for more than five people.
  • Turning on the card switch for convenience and forgetting it. Check the monthly limit under Settings, Billing.

If something was spent that you did not expect

Open Settings, Plans and usage. The activity chart and credit numbers show what ran and when. Disconnect the app if needed, tighten the rules above and write to support@argorant.com if a charge looks wrong.

Related
MCP tools, scopes and spending safetyKey rotation for agents402: payment required, and what to do about itCLI exit codes and what to do with each

Still stuck? support@argorant.com

Help homeargorant.comPricingRoadmap
support@argorant.com