What shipped.
Newest first.

v33 · agents verify proof bundles over MCP

New read-only MCP tool verify_bundle: the open Curbi Proof verifier, unchanged, available to any agent. It runs on a bundle with keys the caller chooses and checks signatures, key windows, RFC 9162 inclusion, attestation binding and history.

v32 · agents can export regressions over MCP

New read-only MCP tool export_regression. It takes exactly policy_check's input, runs the same check on the server, and returns:

• curbi.regression JSON;
• a dependency-free node:test file;
• the command to run it.

v31 · findings you can keep, and a start block for agents

Every finding can now leave the page as something a developer keeps:

• Marked spans: the exact characters are highlighted, with the rule name next to them.
• Copy as regression case: curbi.regression JSON with each reply, its expected verdict and its rules.
• Download runnable test: a dependency-free node:test file that replays the cases against the Check API in one batch (CURBI_API overrides the endpoint).
• Open in Policy Studio: the policy and the failing reply, carried in the URL fragment, so nothing is sent to a server.

v30 · working footer pages

Nine footer pages that were mostly prose now carry a working panel, generated by running the shipped checker, gate, MCP handler and proof verifier (v30/gen.mjs, deterministic). Every command shown is executed against the exported Worker in tests.

v29 · portable evidence and replayable repairs

Curbi’s homepage, footer, product pages and agent manifest now use one direction: check agent action permissions, verify signed records offline, and replay integration failures. The homepage and developer pages share the dark Curbi palette and original multicolor mark. The wordmark uses its existing alpha mask on dark surfaces; the source logo bitmap is unchanged.

v28 · developer layer and hardened footer

Every example on the API and MCP pages is produced by running the real checker, Check API handler and public MCP handler at build time, and every snippet is executed against the exported Worker in the test suite. New: a live JSON-RPC console on /mcp/, MCP Chaos, Six promises, this changelog and a rebuilt connectors page. One data-driven footer is on every page, and the build fails on any broken link.

v27 · developer hub and public MCP preview

The API and MCP now have dedicated entry pages. Explore and Products organize the actual tools; Repair Lab has a source download; platform readiness, workflow and support pages have been rebuilt. The shared navigation and expanded footer cover all 38 HTML pages, removing duplicate footers and obsolete homepage section links.

v26 · simple entry, tested action recovery

Homepage and Action Lab now show a visible reply input, plain-language simulated refund decisions, evidence disclosures and a developer-only footer. Public reply-check API example makes no execution claim. Standalone Node 24 regression download bundles the actual checker and gate; it requires no private checkout or npm install. Existing Studio/auth and other product routes remain available. Published legal summaries are preserved at /legal/.

v23 · build, check, integrate

• Six browser-local synthetic policy challenges; initial4/6 → completed6/6. Invalid JSON reports no run. Mobile390px and desktop1280px inspected without horizontal overflow.
• Public MIT source-only starter ZIP: runner, two policies, six cases, README, LICENSE. No account/key/model needed. License applies only to these examples, not private engine or data. No rewards or earnings program promised.
• Signin/register now route to existing workspace or account-free tools. Existing workspace email/password+Google button observed; authenticated login not tested. Unified GitHub/SSO/passkeys are not connected.
• Discovery advertises the starter and entry limitations. Header CTAs preserved.

v22 · citeable facts and one retrieval catalog

Six direct answers generated from v22/answers.json into visible HTML, matching FAQ structured data, canonical citation anchors, Markdown and /docs/catalog.json. Stable IDs and explicitly scoped Markdown byte hashes give engineering a fixed retrieval source. /evidence/ and /release.json publish bounded synthetic engineering evidence and availability. Canonical links on primary pages; sitemap, manifest, footer and optional llms.txt map the new references. No new popup, marketing statistic or hosted-service claim.

v21 · agent-first discovery and signed amount lab

• /agents/: static, readable agent quickstart with an executable request, error handling and trust limits.
• /ai.json and /.well-known/curbi-ai.json: capability manifest linking the actual OpenAPI and policy schema; explicit unavailable capabilities.
• /llms.txt, sitemap and robots discovery; HTML service-description/capability links.
• /lab/: editable signed amount example, real Ed25519 signing/verification with fresh temporary local keys, original-text findings, failed/unknown outcomes, public evidence inspection and report export. No API upload or provider write.

v19 · write to reply

Adopts Muse's focus on teams building agents that write to systems. Replaces generic homepage samples with three synthetic discount outcomes running through the existing public checker. Highlights exact offending spans; distinguishes permission, attempt and completion. Static initial HTML remains useful before JavaScript loads. Get started / Contact sales, playground, API, pricing and existing navigation remain available.

v18 · policy workbench and versioned contract

engineering input: PR #121 at 6dd2dbe3a8fe6d70cdf5b0276c79ab80a427437d.

v17 · delivery optimization and product navigation

Production · deployment · rollback

v16 · reviewed receipts and visual refinement

GitLab-inspired structure in Curbi colors: white canvas, cyan/blue/magenta spectrum, subtle grid, tighter hero, dark architecture band, compact API example, service disclosure and clearer developer path. No borrowed logos, product screenshots or customer claims. Neutral empty checker state. API browser navigation redirects to docs; machine GET remains JSON. Receipt Lab sends fixed synthetic examples only and compares optional assertion acceptance with required-signature rejection. No action is executed.

v15 · Curbi Policy Platform v15 — deployed 2026-10-27

• Public deterministic Check API, browser playground, developer documentation and shared footer.
• Broad AI-policy homepage, platform readiness page and product-based /pricing/.
• Policy Check and Check API free preview; no paid platform offer invented. Existing Revenue OS terms and checkout remain separate.
• Explicit caller-supplied-unverified records in response metadata and documentation. No provider verification, event binding, ledger resolver or autonomous execution is claimed.