CurbiCurbi

Curbi / Definition · public access

What is Curbi?

Curbi is an independent auditor for agent claims: a developer workbench to check configured replies, verify caller-trusted signed records and keep failures as runnable regression tests.

For developers and AI agents building with APIs, MCP and agent frameworks.

Run a live check →Download the CI starterUse the GitHub template ↗

Can it do that, did it happen, and if it broke, here's the fix.

The public tools check configured rules and supplied records, lints policies, exports runnable tests, and verifies proof bundles with caller-chosen trusted keys.

Curbi is for the boundary between an agent's reply, the rules it must follow, and the evidence supplied for its claims. It works alongside models, agent frameworks and gateways. Independent means a separate check from the system making the claim; it does not mean independent observation of a provider outcome. Curbi does not route model or provider traffic.

Try it. Keep what broke.

01 · CHECKPaste a reply and configure your rules. Inspect the original span and rule.
02 · KEEPDownload that captured reply as a runnable regression.
03 · REPLAYRun node --test locally or in CI. A changed expectation must fail.

Three things to keep distinct

A check
Compares text with configured rules and supplied records. It catches the patterns implemented by those rules.
A verified bundle
Checks signed statements and log proofs against keys the caller explicitly trusts. The signature identifies what the issuer said.
A regression
Keeps one observed behavior as a repeatable test. Review its expectations: a test can preserve a bug.
A passing check means no configured rule fired. It does not prove truth or grant permission. A verified bundle authenticates signed statements and supplied log evidence; it does not independently prove a provider action happened.

For a developer. For an agent.

The HTTP API accepts the same policy and replies your app uses. MCP exposes 6 read-only tools:

HTTP quickstart → · Machine-readable setup → · Current source and release checks →

What is live, and what is not?

Production signing is off. Hosted TEST evidence has its own trust scope. Customer action receipts remain unavailable.

The labs include labeled synthetic fixtures. Published framework adapters are available on npm and PyPI. The separate AI SDK HTTP adapter is available as a versioned first-party download. This is not a production execution gateway.

Evidence and limits · Proof format · Service promises

Published by Curbi, Inc. · Updated October 29, 2026 · Definition JSON · hello@curbi.io

Read a failure you can reproduce.

One extra character turned an approved $49 into $49,000. Our old checker missed it. A failing test, parser fix and permanent regression make the difference inspectable.

The extra k → · Synthetic checker test; no transaction or payment occurred.

Run a check. Keep the result.

Check a reply Connect your agent →

A pass means no configured rule fired. It grants no execution authority or provider-completion proof.

One real provider. One inspectable receipt.

Stripe test mode: one 50-cent refund, a lost readback response, SQLite restart, GET-only recovery, one receipt. The signed bundle includes the supporting observation; tampering is rejected offline.

Locally qualified against the real Stripe test API. Not a hosted execution service, Stripe-signed proof or production signing.

Inspect the run and verify the bundle → Read the ontology →