Start from a policy
that already works.

Use one

Download a file into your repo, then pass its policy to your package. Put it in curbi/ and the Curbi Check Action replays its cases on every pull request.

curl -sO https://curbi.io/policies/saas-billing.json

# Python (curbi-langchain or curbi-llamaindex)
import json
policy = json.load(open("saas-billing.json"))["policy"]
chain = model | StrOutputParser() | curbi_guard(policy=policy)

// JavaScript (@curbi/vercel)
import t from './saas-billing.json' with { type: 'json' };
const checked = wrapLanguageModel({ model, middleware: curbiMiddleware({ policy: t.policy }) });

Edit prices, phrases and actions to match your business, then run its cases again. A template is a starting point: a pass means no configured rule fired, not that a reply is correct.

AI agents

AI agents: verification policy for AI startups

A verification policy, not a security product. For AI startups whose agents act on accounts, data, credentials or billing. Catches an agent claiming a security or data action with no record (deleted, revoked, rotated, patched, disabled, restored, exported, granted, deployed, refunded), security and compliance assurances it must not give, guarantees, and prices or discounts that differ from yours. Deterministic and checked locally. It does not detect prompt injection, leaked secrets, paraphrased claims or attacks, and it does not encrypt, authenticate, patch or audit anything.

Download ai-agent-verification.json · 19 tested cases

Make it yours:

• Map each action to the operation your system records, so a real DONE record lets the claim pass.
• Replace the sample prices with your own; the live checker then flags any discount or price for those items that you have not approved. Remove rules that do not apply.
• Add the exact security and compliance phrases your agents must never say. Phrases are literal: a reworded claim is not caught.

AI agents

AI agents: coding-agent claims

A verification policy, not a security product. For coding agents that commit, push, merge, deploy or run tests. Catches an agent saying it committed, pushed, merged, deployed, released, published, migrated, installed, reverted, opened a pull request, ran tests or fixed something with no record; and claims that tests pass, CI is green or the change is production ready. A fix claim needs a passing-test record. Deterministic and checked locally. It does not run tests, read CI or inspect code, does not check which branch or commit a record names, and a reworded claim is not caught.

Download coding-agent-claims.json · 20 tested cases

Make it yours:

• Map each verb to the operation your CI or git host records (for example a push event or a passing test run), so a real DONE record lets the claim pass.
• Add the exact status phrases your agents use about tests and CI. Phrases are literal.
• Remove verbs your agent never claims, to keep reviews readable.

E-commerce

E-commerce: orders, refunds and pricing

For shopping and order-support agents. Catches refunds, cancellations and shipments claimed with no record, wrong shipping or add-on prices, unapproved discounts, and promises the store does not make.

Download ecommerce-orders.json · 12 tested cases

Make it yours:

• Replace prices with your fees and add-ons. Order totals vary, so unknownAmounts is "allow".
• Map each action to the operation your order system records.
• List your real promotions in discounts, or remove it so every "% off" is flagged.

Finance

Finance: transactions, fees and limits

For banking and payments agents. Catches transfers, reversals, waivers and approvals claimed with no record, wrong fees and limits, and promises about returns, risk or taxes.

Download finance-transactions.json · 12 tested cases

Make it yours:

• Replace fees and limits with yours: each is a "price" the agent must state exactly.
• Customer amounts vary, so unknownAmounts is "allow".
• This checks what the agent says. It does not enforce limits on real transactions.

AI agents

AI agents: MCP and tool-call claims

A verification policy, not a security product. For agents that use MCP servers or other tools. Catches an agent saying it called a tool, or that a tool sent, created, updated, uploaded or booked something, with no record of the call; assurances that a call succeeded; and a tool result restated differently from the value on record. Deterministic and checked locally. It does not see the tool calls themselves, does not check which tool a record came from, does not read tool results unless the host puts them in the record, and does not detect prompt injection in tool output.

Download mcp-tool-claims.json · 18 tested cases

Make it yours:

• Map each verb to the operation your host records for that tool call, so a real DONE record lets the claim pass.
• Put the values a tool returned into record.facts (near, kind, value), written the way the agent writes them ("$40", "3pm"); a different spelling of the same value is flagged.
• Add the success phrases your agents use. Phrases are literal: a reworded claim is not caught.

SaaS

SaaS: billing, plan changes and trials

For billing and account agents. Catches plan changes, refunds, credits and trial extensions claimed with no record, wrong plan prices, unapproved discounts, and promises like unlimited usage or locked prices.

Download saas-billing.json · 13 tested cases

Make it yours:

• Replace the plans and monthly prices with yours; add aliases your customers use.
• Keep discounts to what you actually offer (here: 20% for annual billing).
• Map each action to the operation your billing system records.

Support

Support: escalation and resolution

For support agents. Catches escalations, assignments, fixes and resolutions claimed with no record, response-time promises, and apologies that make commitments.

Download support-escalation.json · 10 tested cases

Make it yours:

• Map each action to the operation your helpdesk records (for example a Zendesk or Linear status change).
• Add the response-time phrases your team must not promise.