What Curbi checks, what the evidence means, and how an agent can use it today.
Curbi checks an AI reply against explicit business rules and supplied action records. It returns structured findings for configured price mismatches, forbidden phrases and unsupported completion claims. The current public access is deterministic, uses no model inference, and does not authorize or execute actions.
Map the action verb to an operation in your policy, then supply its outcome record. Curbi reports action:unrecorded when the reply claims completion without a matching DONE effect. Caller-supplied DONE records are unverified assertions. Put authorization in the execution layer before the write; check the reply after the outcome.
The amount-aware checker can flag a dollar amount that contradicts an amount-bearing signed receipt for the claimed operation in the same sentence. The signed amount lab demonstrates this with temporary local keys and sample data. Those keys are not trusted by the public API, and the lab does not prove a real provider action occurred.
Fetch Curbi’s capability manifest, follow its OpenAPI contract and POST a policy and reply to https://api.curbi.io/v1/check. Inspect pass, typed findings and original-text spans. Public rate-limited checks can omit a key; managed metered access requires an API key. A passing check means no configured issues were found; it is not general truth, compliance certification or permission to act.
Curbi offers six read-only tools at https://api.curbi.io/mcp: policy_check, validate_policy, export_regression, verify_bundle, search and fetch. verify_bundle checks signed statements and supplied log proofs against caller-chosen trusted keys; policy_check does not verify receipts. No private project access, provider calls or action authority. A verified bundle does not prove the provider action happened. Official SDK transport tested; individual client onboarding remains unverified.
Trust the result only within the submitted policy and documented checker behavior. Supplied records are unverified; receipt verification depends on configured keys and its stated scope. Production signing is off in this release. The checker does not persist request bodies, but that does not mean hosting-level request metadata is absent. No SLA is included.
A browser workspace at /sandbox/ for building and breaking a policy: run real Check API calls, see the exact text each finding points to, keep request history, and download the result as a runnable CI regression test. Workspaces live in your browser only; there are no hosted accounts yet.
Call the read-only MCP tool validate_policy with your policy. It applies the same schema rules as policy_check, then warns about likely mistakes: no rules, duplicate or overlapping item names, aliases that drop the item name, duplicate forbid ids or phrases, prices that cannot be stated in cents, and rules the checker never fires on. Nothing is stored.
Use export_regression over MCP, or Download runnable test on /check/ or /sandbox/. Both return the same file: a dependency-free node:test that replays your replies against the Check API. Expected verdicts come from the checker, never from the caller. Run it with node --test.
Call the read-only MCP tool verify_bundle with the bundle and the public keys you trust. It checks signatures, key windows, inclusion in the signed log head (RFC 9162), attestation binding and history consistency, offline. Keys inside the bundle are never trusted. A verified bundle shows what a trusted key signed and logged, not that the provider action happened.
Yes: github.com/Curbi-Inc/curbi-regression-template (MIT). Fork it, edit the policy and cases, push, and CI checks your agent replies against Curbi on every commit. Deterministic, no model calls.
Local checks and six public read-only MCP tools are free. Managed Check API costs $1 per 1,000 checked replies in USD, billed monthly. Each reply in a batch counts separately. The current safety limit is 100,000 replies per key per calendar month, not an account-wide spend cap. No SLA is included. Production action signing remains off.
Read https://curbi.io/status.json. It reports the live release, the source commit, the tool count, whether production signing is on (it is off), and the latest release-check result. The counts are self-reported by the release process, not independent verification.
Add Curbi as a custom connector with the URL https://api.curbi.io/mcp (remote MCP over Streamable HTTP, free rate-limited public tools, no sign-in). In Claude settings, open Connectors, add a custom connector and paste the URL. Claude can then call the six read-only tools: policy_check, validate_policy, export_regression, verify_bundle, search and fetch. Claude Code: claude mcp add --transport http curbi https://api.curbi.io/mcp
Create a connector with the MCP server URL https://api.curbi.io/mcp and no authentication (where ChatGPT offers custom connectors, for example in developer mode). Curbi's search and fetch tools return results in the shape ChatGPT expects (search: id, title, url; fetch: id, title, text, url), and the other four tools are read-only checks. Nothing is stored and no tool can act on your accounts.
Yes. curbi.io/policies/ has tested templates for e-commerce (orders, refunds, shipping prices, discounts), SaaS (plan changes, trials, plan prices, annual discounts), support (escalations, fixes, response-time promises) and finance (transfers, waivers, fees, limits, promises about returns and risk). Each file holds the policy and the replies that show what it catches; put it in curbi/ and the Curbi Check GitHub Action replays them on every pull request.
YOUR NEXT ACTION
Copy this command into a terminal. No signup or key. It sends a synthetic reply and policy to Curbi and returns three findings: action:unrecorded, price:mismatch and forbid:guarantee.
curl -fsS https://api.curbi.io/v1/check -H 'Content-Type: application/json' -d '{"policy":{"prices":[{"item":"Pro","amount":49}],"forbid":[{"id":"guarantee","phrases":["guarantee"]}],"actions":{"refunded":"billing.refund"}},"reply":"Done! I'\''ve refunded you, Pro is just $29, and we guarantee it."}'What next: edit the policy in your sandbox, then export a runnable regression from the MCP console. Connect your framework → · All commands as JSON →
A pass means no configured rule fired. It grants no execution authority or provider-completion proof. public access; production signing is off.