What shipped, newest first.
The client review, the verifier, the gateway and the site, from September 2026. Dates are the day a change reached a user; versions identify the released package where one is listed.
2026
0.1.4Improvement The Claude extension guides your first month and keeps the exact report.
The local extension now starts with a short result and keeps the full report available. For a real month, its setup checklist shows the files and dates it has, contract terms to confirm and the Fin outcomes total it needs from the custom report. The result counts every finding category before highlighting the main items. Choose separate folders for read-only imports, case history and reports in the extension settings; the bundled demo needs no folder setup. Existing case history and confirmed terms are retained when you update. One local review count covers one vendor billing account for one billing period, including multiple invoices, corrections, reruns and follow-up. This free preview makes no charge and sends no usage to ScopeBlind. The invoice engine is unchanged. Built for Mac and Windows, tested so far on Mac. ScopeBlind receives no files; documents attached to Claude and tool results are handled by Claude. Home, About and llms.txt now describe checks both sides can run without calling ScopeBlind’s own work independent.
Improvement The checker dates a missing deduction, and the free checks come to Claude Desktop.
When an invoice does not show a deduction that Fin’s rule promises, the checker now gives the date to raise it with Fin’s Billing Ops team before you pay. That date is Fin’s request, not a legal deadline. Where it is too early to tell whether an earlier charge is owed back, the result named your contract’s price as the amount charged; it now counts the outcomes charged and values them at that price. And the same free checks now come as an extension for Claude Desktop, tested so far on Mac: download it from the bill check page, open it in Claude Desktop, and ask Claude to run the demo or check your month. Its tools read the files at the paths Claude gives them and upload nothing to ScopeBlind.
0.32.4Fix Every protect-mcp command in Docs and the setup pages installs 0.32.4, a security release.
protect-mcp 0.32.4 fixes the advisory published for it on 25 October 2026 (GHSA-xf4c-59j2-hgf2). In 0.32.3 and earlier, a flag written as --flag=value was ignored: a gateway or hook server started with --cedar=./rules ran without that policy and allowed calls the rules forbid, and evaluate --format=claude evaluated an empty tool name. Every command now reads --flag=value as --flag value, refuses a value given to a switch, and denies a call that names no tool before any policy is consulted. The Claude Code plugin relay in protect-mcp’s repository, which is not part of the npm package, ran no check when it was started through a path with a symbolic link, and the Cursor preview’s relay did the same for a path that differed only in letter case; both now always run their check and still refuse a call when the gate cannot be reached. 0.32.4 also checks a host’s hook payload before deciding or recording anything, writes output larger than a pipe holds whole, and writes every receipt chain link in the sha256: form. Update with npm install protect-mcp@0.32.4, and reinstall the Claude Code or Cursor plugin if you use one. The protect-mcp commands in Docs and on the gateway and repository setup pages, the receiver that repository setup points to, and the files under /releases now name 0.32.4, where they named 0.32.3.
Fix /status and the protect-mcp first run name @veritasacta/verify 0.11.1.
@veritasacta/verify 0.11.1, released on 25 October 2026, writes output larger than a pipe holds whole. Through a shell pipe, on macOS and on Linux, a result over 64 KiB was cut short while the exit status still gave the outcome: verify --replay-chain with --json piped into jq read 65,536 bytes of a 193,133-byte report, with exit 0, and jq failed on the broken JSON, while the same report written to a file was whole. Each run now writes the same output and messages and gives the same exit status as before, and all of the output arrives; the verifier’s test suite also runs on fixtures of its own. /status names 0.11.1, and the first run on /protect-mcp installs it.
Fix Verified runs with Codex work with protect-mcp 0.32.4.
From 0.32.4, protect-mcp prints {} for a Codex hook, since Codex rejects the fields it used to print there. The hook this harness writes for each task read the decision from that output, so with 0.32.4 it would have refused every Codex tool call, allowed ones included. It now reads the decision and the request id from the receipt the call appended, and the site’s tests run an allowed and a refused call through that hook for Claude Code and Codex. The published runs were made before 0.32.4 and are unaffected.
Feature A new page reports protect-mcp on two public agent benchmarks.
/benchmarks reports replays of the runs AgentDojo and tau2-bench publish, passed call by call through protect-mcp 0.32.2’s Cedar gate under rules fixed before each replay, with no model called. On AgentDojo, the destinations-only part A denied the first attacker call in 76.6% of attack traces, and 97.5% of successful honest runs had no denial; the full policy, whose B and C rules were written after reading the attacks, gives 98.7% and 97.2%. On tau2-bench retail, 52 of 52 planted failures were caught, and 33 runs that the answer key accepts would not have reached the gold state under enforcement. The page says what the replays do not show, that this is not an independent evaluation, and that ScopeBlind did the work with AI coding agents. The protect-mcp page links to it.
0.32.3Fix Every protect-mcp command in Docs and the setup pages installs 0.32.3, with the 0.32.2 security fix.
protect-mcp 0.32.2 fixes the advisory published for it on 25 October 2026 (GHSA-2mv7-2jp4-p55f). ConfidentialGate, a beta part of its library, granted the privileged tier from attestation documents it never verified; it now refuses every document unless the code using it supplies a verifier. verifyEvidenceAttestation counted oracle and witness attestations as valid on a flag their producer sets, and no longer does. protect-mcp’s own gateway, hooks and receipts call neither. In a Claude Code PostToolUse hook, sign --format claude printed a “decision” of “allow”, which Claude Code rejects there, so every call ended with a hook error; it now prints the receipt’s decision as receipt_decision, and the other formats keep decision. Its optional Cedar engine now starts at 4.13, past GHSA-m84h-6f26-p3p7. 0.32.3, released the same day, makes hook setup reversible: init-hooks --dry-run shows the proposed changes without writing files, an installation keeps unrelated settings, hooks, policies and keys while a backup and an ownership manifest record the entries it adds, and init-hooks --uninstall removes only the unchanged entries it added. The new doctor command checks the directory, configuration, Cedar policies and signing key the runtime uses and makes a local sign and verify check, without inferring a host connection or active enforcement from the files it finds. Generated MCP wrapper configurations pin the running version rather than @latest, and offboarding an older installation edits the project that --dir names. The protect-mcp commands in Docs and on the gateway and repository setup pages, the receiver that repository setup points to, and the files under /releases now name 0.32.3, where they named 0.32.1.
Feature protect-mcp has its own page.
/protect-mcp, which led to the gateway section of Docs, is now a page for developers: a first run that installs protect-mcp 0.32.3 and the verifier, then evaluates an allowed file read and a denied force-push without executing either; the three ways to connect an agent (Claude Code hooks, an MCP server wrapper, other coding agents); a captured demo run; and how to contribute. The footer’s other tools and the About page link to it.
Fix The Cedar engine in ScopeBlind’s own code moves to 4.14.0, past an advisory.
ScopeBlind’s trading mandate engine, and the Cedar engine file the site serves, used @cedar-policy/cedar-wasm 4.10.0, inside GHSA-m84h-6f26-p3p7: in Cedar 4.3 to 4.12, parsing a has expression with a nested attribute path could exhaust memory. Both now use 4.14.0. The policies, their digests and the mandate digest are unchanged.
Fix Verified runs with Claude Code work with protect-mcp 0.32.2 and later.
protect-mcp 0.32.2 prints a Claude Code hook’s receipt decision as receipt_decision. The hook this harness writes for each task read only decision, so with 0.32.2 or later it would have refused every Claude Code tool call, allowed ones included. It now reads receipt_decision and falls back to decision, and the site’s tests run an allowed and a refused call through that hook for Claude Code and Codex. The published runs were made before 0.32.2 and are unaffected.
0.32.1Fix Every protect-mcp command in Docs and the setup pages installs 0.32.1, a security release.
protect-mcp 0.32.0 fixed the three advisories published for it on 23 October 2026: its --http gateway accepted unauthenticated tool calls on every network interface, an agent could approve its own calls and some policy rules were not applied, and its local servers accepted requests from any web page. 0.32.1, released a day later, moves the dependencies that arrive with the MCP SDK past their advisories, among them body-parser and zod, and an enforcing gate now refuses a call whose receipt it cannot record and stops at startup on a receipt log whose chain it cannot continue. The protect-mcp commands in Docs and on the gateway and repository setup pages, the receiver that repository setup points to, and the files under /releases now name 0.32.1, where they named 0.31.1. From 0.32.0 --http listens on 127.0.0.1 and needs a bearer token, so the Docs command that serves an agent profile at a local address now sets one and says to give it to the connector.
Fix Verify reads a protect-mcp receipt exactly as protect-mcp signs it.
Verify read the signed bytes of a protect-mcp receipt in an order that is not RFC 8785. A genuine receipt whose payload had integer-like keys, such as “10” and “9”, showed as invalid, and a receipt with an unsigned “__proto__” member added showed as valid. It now reads them as protect-mcp 0.32.1 and @veritasacta/verify 0.11.0 do: a signature made in the old order is named and not counted valid, and a log that names a member twice is not read. The parser that Verify and the hosted connector use for other records also kept a “__proto__” member out of sight of their checks. It now reads it as a member, so a signature has to cover it.
Fix The commands that check a record give the signer’s key, as @veritasacta/verify 0.11.0 requires.
@veritasacta/verify 0.11.0, a security release, calls a result that rests only on a key the file carries undecidable (exit 2), where earlier versions exited 0: repository evidence, proof packs, standards and run manifests among them. The commands on How it works, Docs, Tools and Verify now give the signer’s key with --key, or for a published run the maintainer key with --maintainer-key, and a receipt log is checked with --replay-chain, without which the verifier did not read one. The evidence pack on Verify signs a digest of its bytes, which the verifier does not read, so Verify now points to the steps beside it instead of a command.
Fix The hosted connector says what it checks.
Docs said the hosted connector’s coordination.verify_record uses the same parser and verifiers as Verify. It reads a record with Verify’s parser, and checks shared task, negotiation, rehearsal and repository evidence, and single records the service signed, against scopeblind.com’s own key. It now also checks protect-mcp gateway receipts, as self-signed records: a receipt carries the key that signed it, so it is never reported as verified, and it means something only once that key is pinned as the gateway operator’s own. Verify reads more formats. The connector’s own description of the tool says the same.
Fix The trading mandate engine blocks an order when one of its rules errors.
Cedar skips a rule whose evaluation fails, for example one that reads an attribute the order does not carry, or a misspelt one. In the engine that holds an agent’s orders to a trading mandate, a forbid rule that failed that way was skipped, and the permit let the order through. The engine now blocks the order with evaluation_error, as protect-mcp has done since 0.7.0. It also names the rules that decided from Cedar’s own reading of the policy, so a hand-written rule with a record or an unless clause keeps its name in the receipt. The shipped and generated mandates decide as before.
Improvement The home page leads with what each resolution a customer confirmed cost.
The sample on the home page now opens with the price per resolution a customer confirmed: in its first month, $4.09, because a customer confirmed 302 of the 1,247 resolutions in Fin’s export and the others were assumed. The difference from the contract, $51.36, follows with its arithmetic. Beside it, what Fin counts as confirmed, and that an assumed resolution is not a failed one: the figure is a question to take into the renewal.
Fix Stripe is listed among the companies that handle data, for the Assisted review’s fee.
The Terms say the Assisted review is paid once, through Stripe, by a payment link or an invoice. Security and Privacy now list Stripe among the companies that process data: it receives the billing contact and the payment, and card details go to Stripe, never to ScopeBlind. The checker itself still sends nothing to anyone.
Fix The renewal scenarios name the invoice’s count as billed resolutions.
Scenario B starts from the resolutions your invoice billed, which can differ by a few from the resolutions in Fin’s export beside it (1,159 and 1,160 in the sample’s second month). The row now says “billed resolutions”, so the two are not read as one number that disagrees.
Feature The blog is back, with the earlier posts kept for the record.
The blog is back. Five of the posts written for the earlier products are kept for the record on one page, each under its old address and each with a note saying what was corrected or removed since; the others are not kept. About now lists the lesson Tom Farley wrote for Microsoft’s open-source course AI Agents for Beginners.
Fix Early access, said as it stands, and a record kept when it cannot be downloaded.
The checker now says where it stands: it has read a real Fin export from ScopeBlind’s own Intercom workspace, but not yet a customer’s month. Before, it said it had been tested on simulated data only. And “Download my record, then start again” now starts again only once your record has reached your browser: if the record cannot be made, nothing is cleared and the page says why.
Fix The checker reads Fin’s export as Intercom writes it.
Intercom names each time column of Fin’s Conversations export with your workspace’s time zone, such as “Conversation created at (Australia/Sydney)”, and writes its times without one. The checker asked which column held the creation time, and read a column chosen there as UTC, hours from your workspace’s time. It now reads each time in the zone its column names, summer time included, so the first answer comes with no question. Found by running Fin on ScopeBlind’s own Intercom workspace and checking its export.
Fix An invoice changed after it was read is read again, every dropped file is accounted for, and Fin Voice lines are read apart.
If you change an invoice after the checker read it, the reading is out of date: the page says so and reads the invoice again before anything can be confirmed, so what is stored is the invoice as it is now. Start again now asks first while your work is kept nowhere (download your record, start again without it, or cancel), and it clears the pasted invoice, the typed lines and the order form too. When you drop several files at once, Fin’s export is read as the export and the first invoice as the invoice, and the page says where each went and names any it did not read. Renewal scenario B now starts from your invoice’s billed resolutions at your contract’s price, and says what its break-even divides. An invoice line naming Fin Voice outcomes is read as its own product, never priced as a support outcome: the result names it among what could not be checked. The sample now opens with the charges, what the record follows next and the renewal, and the home page says that getting Fin’s export can take longer than the check.
Improvement One email for each purpose, a credit email that asks for the credit note, and Fin for Sales lines read apart.
The bill check now drafts one email for each purpose: your questions about an invoice, more on them when Fin asks to see more, and the credit Fin agreed, which asks for the credit note and where it was applied, since a standard credit note is applied to the invoice it is issued for. The emails carry no internal reference numbers and no line about not disputing the invoice; beside the email, that line stays for you. An invoice line naming Fin for Sales outcomes (qualifications at $9.99, disqualifications and self-serve routings at $0.99) is read as its own product, never priced as a support outcome. And "Copy for your AI assistant" now ends with where each of your questions for Fin stands, and its next step.
Improvement The paid review leads with what you get, and renewal scenarios run on your own figures.
Assisted review’s card now leads with what you get, its terms beneath, and says when each part comes: the packet within five business days of complete files, then the invoice, then up to two rounds of replies over the 30 days after delivery, and, within 60 days of delivery, confirmation that a credit Fin agreed was issued and where it was applied, plus one check of your next invoice for the deductions Fin’s rules promise, which the terms now state. A standard credit note is applied to the invoice it is issued for, and only a credit to your account balance shows on a later invoice, so the checker no longer sends you to the next invoice for a credit Fin agreed: it asks for the credit note and where it was applied. Where the sample shows $0.99 against $0.95, it now says its order form has no committed amount, since Intercom bills usage over a contracted amount at list price. “Is this right for us?” now answers a small bill with a renewal close, a case open or several months honestly, and every email it offers carries your answers; beside each email link, Copy address and Copy email help on webmail. The sample packet and the checker’s renewal figures add two Fin contract scenarios on your own volume, a lower rate for a committed volume and billing only resolutions a customer confirmed, each with its arithmetic and the decision it supports, computed on the page and labelled as scenarios, not quotes or findings. In the checker, leaving now asks first while terms are typed and not applied, a file is read and not yet checked, or an edited email is not copied, and a record opened again leads the page. The billing rules keep their index in view on a wide screen.
Improvement One buying page, a clearer header, and the checker as a task of its own.
Pricing and Assisted review are now one page: the free checker and Assisted review side by side, then everything the review hands over, and the old Pricing address leads there. The header reads Try a sample, Pricing, Help and About, with "Check your bill, free". The checker has a header of its own: the logo, "Bill check" and Help, which opens in a new tab, and it asks before you leave while work of your own is neither downloaded nor remembered. Six marketing pages (the home page, Pricing, the sample packet, the Fin billing rules, Help and About) now count visits with Cloudflare Web Analytics, which sets no cookies. It is never on the checker, or on any page that reads your files.
Feature Assisted review is now one finished job, with a sample packet to see it.
Assisted review now reconciles your Fin bills, prepares your questions and follows one case through: the reconciliation, on a screen share by default so your files stay on your computer; a findings brief reviewed by Tom Farley; a case ready for you to send; up to two rounds of replies within 30 days, which we draft and you send; one check of the next invoice; and a finance and renewal brief. It covers one contract and up to three consecutive months, for US$500 fixed, the design-partner price for the first three companies, invoiced after delivery. If the checks cannot run on your files, there is nothing to pay. The sample packet shows every part on a fictional company, ready to print, and four quick questions on the page say whether the review fits, or whether the free checker is enough. The terms, Pricing, llms.txt and the commercial manifest say the same.
Improvement Ready for your AI assistant, and the example review in brief.
llms.txt, the file AI assistants read about a site, now says what ScopeBlind is, who the bill check is for, the questions it answers and what it cannot check, and links the ten billing rules, the example review, Help and Pricing. It asks assistants never to have a person paste a Fin export or customer conversations into a chat, and to send them to the checker, which runs in their own browser. The ten rules and the example review each have "Copy for your AI assistant": the page as text, the rules with their sources and dates and the example labelled fictional, ending with that same request. The example review opens on its summary: what is still open, the top finding, the renewal figure, the date to keep and what the review is for at renewal, with the full review a click away. About says what ScopeBlind builds: checks for the work AI agents do between companies, starting with Fin bills.
Fix After a reply, the whole result follows it, and a record opened again says where it stands.
In the bill check, recording that you sent a question, Fin's reply, a credit note or a correction now updates every part of the result from your record: the next step, the month's summary, the questions, the dates to keep and the email to Fin. If you edited the email, your version stays, and the updated draft is offered beside it until you choose. Opening last month's record shows at once the last invoice checked, the questions put to Fin, with any that need evidence first counted apart, and the credits you expect, with buttons to record a reply, add a credit or check the next invoice, and no new check needed. The month's summary no longer asks you to send Fin a question there is no email for: one that needs evidence first is counted on its own. The email's subject is a field of its own, with its own copy button, and a short email opens in your email app. Fin in Zendesk now starts with Zendesk's ticket export too: how many of the tickets Fin tagged as resolutions carry fin-hard-resolution, the tag for a resolution the customer confirmed, with the export named as Zendesk's at every step. The calendar reminder says which export to bring, and the affiliation line names Zendesk.
Improvement Each question keeps its name, and a record made by mistake can be corrected.
In the bill check, each question you put to Fin keeps its own name, period, invoice and amount whatever you record of it, and where it stands is a line of its own. A reply or a date recorded by mistake can be corrected: the correction is a new entry in your record, the first is kept, and the check says what was corrected and when. "Download your record" always downloads the record as it is now, and the page says when you last downloaded it and whether it has changed since. On a phone, the follow-up buttons are full size. The email to Fin opens for you to read before you copy it, and you can edit anything in it: your record notes which questions you sent, not the wording. When you pay Fin's published prices, the email says it was checked against them, not against a contract. Fin's date for raising a missing credit is said once, under the dates to keep; the sample shows its result first; the first screen is lighter; and the note for us about a file the checker could not read says what it holds, and can be edited before you send it.
Improvement Help for the bill check, and what the paid review hands over.
Docs in the navigation is now Help: one page for checking a Fin bill, with what you need, how to get Fin's export from Intercom or Zendesk, the outcomes total, how your record carries to next month, the formats the checker reads and the checks it runs, each tested on simulated data so far, and what to do if a file is not read. Docs is the technical reference, and its old addresses still work. Assisted review shows an illustrative review of a fictional company, every figure from the checker's own sample run. The home's sample is now the difference, its calculation and the next step, and the limits are one short row.
Feature What you asked Fin, and what happened, kept with your check.
Each check now keeps a short table of the questions in its email to Fin: the amount, where each stands, and buttons to record that you sent it, what Fin replied (with its ticket, if any) and a credit note when one arrives. Each entry is your own record, said as yours, and next month's check reads it beside the next invoice. When you open your record again, the result leads with what changed since your last check. The record section says plainly whether your record is kept in this browser, and when, or that it is not kept anywhere yet.
Improvement One thing at a time on the first screen, and the right file read where it belongs.
The bill check opens on Fin's export as step 01, as the home page does, and the steps below it wait until the export is read or you choose to start with the invoice, so one drop zone is open at a time. An invoice dropped where the export goes is read as the invoice, and an export dropped where the invoice goes is read as the export, each with a line saying where it went. The help for Fin's export carries a request, ready to copy, for the two Intercom permissions it needs. If the checker cannot read an export, it offers a note for us with the file's name, size, column names and number of lines, never a row, which you send only if you choose.
Fix The email to Fin says only what was checked.
The bill check's email to Fin now names only what the check read: an invoice checked on its own is "checked against our contract", with no claim about Fin's conversation list, and Fin in Zendesk is checked against the tags Fin put on your own Zendesk tickets. Each value in your terms says how it got there: from your order form only where the order form reader read it, entered by you, or confirmed by you. Looking at the sample no longer clears your own check, a printed result carries its calculation, what could not be checked and the dates to keep, and the sample's second month follows the ending you pick at once.
Feature Fin in Zendesk, from Zendesk's own export.
The Fin bill check has a setup for Fin in Zendesk, counted from Zendesk's own ticket export by the tags Fin puts on each ticket it resolves, with how to get that export from Zendesk's Admin Center. Its one question for your terms names Fin's plan for an existing helpdesk, as Fin's pricing page lists it: a fixed fee of $49.00 a month, with 50 outcomes included.
Improvement One page to print, three folds, plainer words.
A result prints on its own, as a page to forward or attach: its date, the figures, what to ask Fin, and the engine it ran on. Its folds are three, not eight: what to send and the dates to keep, the renewal figures, and how it was checked. And the checker's first lines say "shows nothing your contract rules out" and "too early to tell", where they said "has no established conflict" and "cannot be told yet".
Improvement Getting Fin's export, step by step.
The export step says how to get Fin's export: where it is in Intercom, the two permissions it needs, the dates to choose, and that above 10,000 rows it comes by email, up to an hour later. With no export yet, a check can start with the invoice, whose prices are checked at once, and take the export when it arrives. On a phone, the checker offers the sample there, or an email to yourself with the page's address.
Improvement The checker runs the engine file you can hash.
The checker page no longer carries its own copy of the engine. It runs /bill-check/engine.js, the file served beside it, pinned by that file's SHA-256, so your browser refuses any other bytes, and the file you can download and hash is the one that ran. The page is also about half the size, and the engine is cached between visits.
Fix The renewal figure, only where the records give it.
The price per resolution a customer confirmed could fall below zero under a plan that includes outcomes, or when the invoice and Fin's export counted differently. Under a plan, its fee now counts and is shared with procedure handoffs at your contract's prices. Where the invoice and the export count differently, or a commitment pays across the term, the checker gives the confirmed count and says why it gives no price.
Feature The order form, read for its terms.
The Fin bill check reads your order form on the terms step: its PDF, read in your browser, or its pasted text. Each term it finds is shown with the line it came from, and goes into the form only when you say it matches. A term the order form does not give plainly is left for you to enter, with why. The order form is not kept.
Improvement Plainer results.
What would settle an open question is one line beneath it, with the detail a click away, and the checker's early-access notice is one line. A result can be copied for your own AI assistant without customer text, and a reminder for next month's check is made on the page as a calendar file.
Feature Fin's export first, and ten of Fin's billing rules.
The Fin bill check can start from Fin's own conversation export, and shows how many resolutions it records as confirmed before any invoice. One question sets your terms when you pay Fin's published prices. Ten of Fin's billing rules, each in Fin's or Intercom's own words, are on a page of their own.
0.31.1Fix The MCP SDK inside protect-mcp, past its OAuth advisory.
protect-mcp 0.31.1 bundles the MCP TypeScript SDK at 1.32.1. The advisory GHSA-6qxp-vccf-f47h is in the SDK’s OAuth client, which protect-mcp neither bundles nor calls, so 0.31.0 was not affected; the release clears the dependency audit this site’s release gate runs. The Claude Code hooks file in the package now pins 0.31.1, where it pinned 0.13.4, and every command on this site names 0.31.1.
Launch The Fin bill check.
A free checker for Fin bills that runs in your browser: it compares each invoice with the terms you enter and Fin's own export, says what differs and what could not be checked, and drafts the questions for you to send. Nothing is uploaded. Assisted review, a person working through up to three months with you, is beside it.
0.31.0Feature A starter policy, and denials that say why.
protect-mcp now ships a starter policy for coding agents: every tool call is allowed except seven named rules, which block a recursive delete of / or the home folder, a force-push to main, reading a credential file, a download piped into a shell, and writes to shell profiles, credential files and the agent's own settings. One command writes it and the key that signs receipts: npx protect-mcp@0.31.0 init --starter. Every deny now names the rule that matched and why, and the hook prints it where the agent reads it, so the agent can change its plan. A tool call carrying a fraction or a null, as browser tools send, is no longer refused under every policy. The shell rules match the command text, so they are a guardrail for an agent working in good faith, not a sandbox.
Improvement Plans, payment and privacy, in writing.
The Terms now say how paid plans work: invoices go out monthly in advance in US dollars, are paid by bank transfer within 14 days, and prices exclude taxes. Team continues after its fourteen free days and is invoiced monthly unless you cancel by email before they end, and we send a reminder at least three days before. Cancelling takes effect at the end of the month already invoiced, which stays payable; cancelling during the free days means nothing is invoiced; there is no refund for part months, and your consumer-law rights are unchanged. If we end a paid plan ourselves, we refund the unused part of the month, and a significant change to the terms comes with 30 days' notice. Founding-team pricing is held for twelve months, and any later price change comes with 30 days' notice. The service is for work, for people aged 18 or over, and ScopeBlind does not train models on your content. The Privacy page now promises to acknowledge a data request within five business days and complete it within 30, and says how long each kind of operational data is kept. Both pages name the company by its ACN and ABN and give one address, support@scopeblind.com, which security.txt also lists.
Fix Enquiries arrive, and data that only runs the service is cleaned up.
Team and design partner enquiries were saved but never emailed to us, because the notification had no subject line and our mail provider refused it; nobody had sent one yet. Data that only runs the service is now removed by a scheduled cleanup every fifteen minutes: hourly counters for the model features after a day, the hashed network address kept for a button demo, a repository setup or a managed trial after 30 days, device link requests nobody approved, the record of which linked device acted in an invoice trial after 30 days, the reviewer name and note an agent suggested for a trial once the trial is over, and sign-in requests that never connected. Signed records people rely on are never removed this way. A published standard page now stays for 365 days from its latest write, as the Privacy page says, instead of lapsing a year after it was published. A new page lets you confirm a request about your data from the browser that holds your key. Retired code that called services the Privacy page does not list has been removed.
Fix The hosted connector works in Claude.
Three things stood between the hosted connector and Claude on the web and in the desktop app, which call it from Anthropic's servers. The connector now accepts requests that name claude.ai as their origin, and still refuses any other site. It lists its tools with an underscore in place of the dot, such as coordination_inspect_workspace, because Claude refuses a dot in a tool name; the dotted names still work for platforms that use them. And it answers ordinary requests that name a newer protocol version, while still refusing the newer handshake itself so the platform falls back to the one it shares. The sign-in metadata now also links the Terms and the Privacy page.
Fix Your agent can find every review it may read.
A hosted connector used to list only the reviews created from its own drafts, so it could not say where a review stood if the organizer had created it directly, such as the first review in a trial. It now also lists every other review in the project that its agreed limits cover: the same two people, repository, branch, files and checks the service already checks on every read. Nothing it can read changed. The connection page also starts over if a platform sends you to a different connection request in the same tab, so it never shows an earlier request than the one its address names.
Feature Your agent can start a trial with you.
Someone with no project and no repository can now get from inside their agent to a working two-person review. When the agent platform connects, choose a new trial on ScopeBlind's demo repository. Ask the agent to start one and it gives you a link, which carries no secret and opens only in the browser where you connected it. There you start the trial and invite your reviewer yourself; the agent never sees the invitation. When the trial's project exists, the agent's connector joins it, and it can prepare and read work once you propose limits naming it and your reviewer agrees them. A connector's key still never signs as a person, before or after it joins. The shared trial's capacity check also stops counting ScopeBlind's own test runs, which the trial limits already ignored.
Feature Connect an agent platform in one step.
An agent platform that supports signing in to an MCP server can now connect to a project without anyone copying a token. Add https://scopeblind.com/mcp and choose connect: the platform opens scopeblind.com in your browser, you pick one of the projects that browser organizes, check the address the platform will return to, and choose Connect. Your browser signs that approval with your own project key; the platform then exchanges a one-time code for its token, straight from ScopeBlind. It is OAuth 2.1 with PKCE for public clients, with client ID metadata documents and a registration that stores nothing. The connection is an ordinary hosted connector: it grants nothing until the organizer proposes limits naming its key and the reviewer agrees them, it appears on the project page, and the organizer can revoke it there. The approval works only in the browser that holds the project; there is no approval by link.
Improvement Your agent can say who is waiting.
Ask your agent where a review stands and it now answers in a few lines, through the hosted connector: who is waiting on whom and for what, what is left, how agents assessed each success criterion, the feedback on this version, and any open Resolve this options with the one decision each needs. It follows the same rules as the project page, so an expired snapshot or an expired approval is reported as waiting on the receiver, not on a person. The agent can read Resolve this options but never choose one; a choice counts only when the named person's own signed feedback records it. Separately, a hosted connector's key is now refused for anything a person does, such as joining a project, owning or reviewing a task, approving or starting a trial, both by the service and by the data itself.
Improvement The hosted connector finds its own project, and can check a record.
A cloud agent no longer needs anyone to paste IDs: the hosted connector belongs to one project, so it finds the limits its two people agreed for it by itself, and lists the reviews created from its drafts. When something is missing, it now says what, in a sentence the agent can pass on: no limits yet, the reviewer has not agreed them, they expired or were used up. A new read-only tool checks a ScopeBlind record someone was sent, exactly as received, with the same verifiers as the Verify page, pinned to scopeblind.com’s own service key; a record that is consistent only with a key it names itself, or that carries text outside its signatures, is never called genuine. The docs no longer say the agent can see open Resolve this options: only the two people can.
Feature A hosted connector for cloud agents.
A cloud agent that cannot run a local command can now reach a project through one address, https://scopeblind.com/mcp, with a token the project's organizer creates on the project page. The token is shown once. Of the token, ScopeBlind keeps only a hash, beside the connector's public key, and it derives the connector's signing key from the token on each request, so nothing stored here can sign as the connector. It offers six of the local agent profile's tools under the same limits and the same human steps: the organizer proposes preparation limits naming the connector's key, the reviewer agrees them, the agent prepares and reports, and people approve, accept and withdraw on their own devices. Revoking a connector stops its token at once, and the service and its database refuse the connector's key from then on. The site also has plain Privacy and Terms pages now, written from what the service actually stores.
Fix Receiver setup, said plainly.
Three things a first rehearsal of the personal-agent loop tripped over. The advanced connection import now says which file it takes (the connection.json that repository setup writes next to receiver-key.json) and names the wrong file when you pick the signed connection record, the receiver configuration or a private key. The connected-receiver button no longer waits for a poll that only ran while the tab was in front: its first check runs in a background tab too, and the page says why the button is waiting and when updates are paused. The receiver command on the task page no longer invents a key path; it says where repository setup and repository connect actually keep the private key, and the docs say the same.
Feature Bring your personal agent.
A personal agent such as Meta's Muse can now prepare shared work here and follow it through: two commands in the agent's own computer give it a ScopeBlind profile served as a local MCP address, which Muse takes as a custom connector. The agent prepares briefs and revisions and reads status, the next decision and the result; every approval, acceptance and withdrawal is still signed by a person on their own device, the receiver still holds the only credential that can change the destination, and the record still leaves with whoever wants it. The docs carry the setup and a public skill says it in the agent's terms.
0.30.0Feature The grant an agent is acting under travels with the call.
An agent platform can now send, with each tool call, the grant its agent is acting under: who the platform is, which agent, what the person allowed, and for how long. The gateway records it on the decision receipt exactly as presented, together with what it was able to check about the platform's signature: unsigned, unverifiable when it holds no key for that platform, valid, or invalid. Naming a platform key in the policy file is what turns a signature into a checked one, so an unchecked context is never reported as a checked one. The open verifier reads the same member and refuses a receipt whose recorded context is not what the platform signed. This follows a new Veritas Acta draft, with seven shared test vectors, three of which a conformant verifier must refuse.
Feature Resolve this.
When a review is stuck, either person can ask for a resolution. The model proposes a few options inside the agreed scope; the service refuses any that leave the allowed files, name a criterion the brief does not have, or need more than one further decision, and hands the refusals back for one more round at most. What comes back is two to four feasible options, each naming the one decision a person still has to make. Choosing one records ordinary signed feedback for the next revision. Three resolutions per task; nothing is approved or applied by any of it.
0.29.0Feature The result you saw is the result delivered.
A published preview is now rendered from bytes your browser fetched and checked, file by file, against the digest in the worker's signed plan; nothing else can load inside it. Every review can be kept as one file: the review package carries the signed evidence, the verified preview bytes and an offline viewer that re-checks all of it on open with no network. A coding result now names the workflow run and attempt that produced it, signed by the worker from inside that run. And once the recipient accepts a result, its preview is served at one address that will not change, resolving to the same verified bytes.
0.28.0Improvement One standard, two enforcers, one answer.
A standard is enforced twice: by the gateway on the agent's machine and by this site, which compiles it, rehearses call logs against it, and reads the receipts. A shared file of decision vectors, six standards and fifty calls with the answer each standard's own words require, now runs in both release gates, so the two cannot drift apart unnoticed. Three readings were aligned on the gateway side: a standard that states neither tools nor a limit admits nothing; a payment standard admits only the tool it names; amounts are read as integer minor units in the standard's currency, spelled exactly. The demo's inspection now asks for itself as soon as both people are present, from the page both of them are actually polling.
0.27.0Feature Start a revision with what changed.
Each success criterion can now say what it relies on: particular files, named checks, or the preview. When a revision arrives, the review opens with the criteria whose wording was revised, then the criteria whose linked files or preview changed in this version, and folds away the criteria whose wording and evidence are both unchanged. A criterion with no links is treated as relying on the whole change, so nothing is quietly carried over. The new version still needs both people’s fresh decision. Also: once both people have joined a shared demo, the inspection is requested for them instead of waiting for a click.
0.26.0Feature Approve on Friday. Deliver on Monday.
An approval now holds until the task’s deadline, up to seven days, instead of fifteen minutes. It stays bound to one exact version and the brief it was made against. Either person can withdraw their own approval until delivery begins, and the withdrawal stays in the record. At delivery the receiver asks for a short-lived permission that is issued only after rechecking both decisions, membership, the destination branch, and the recorded evidence, then applies within two minutes or not at all. A revised version needs its own two decisions and nothing else. Receivers on protect-mcp 0.26.0 or later apply durable approvals; the demo repository is updated with this release.
Improvement The queue keeps moving when both tabs are in the background.
Review, demo, and trial pages now keep checking for the other person about once a minute while their tab is hidden, so a requested inspection or apply is woken even if both people switched away; in front, they still update every five seconds. A busy model provider no longer ends a negotiation: the step is retried with a growing pause. Internal release checks now run in their own allowance, so a deploy never spends a visitor’s demo or trial budget.
Improvement Five strangers walked the site. Ten things got fixed.
A simulated stranger test of the live site, two of them running a real review in separate browsers, turned up ten things worth fixing. The home page now leads with the friend trial and says what a client review needs. The AI trial tells you the real reason when it cannot start and no longer calls a rejected request saved. Review pages say when they last checked for the other person, refresh the moment you return to the tab, and say so when an inspection has not arrived. The person opening an invitation is greeted as the invited reviewer. Buttons that wait for a tick tell you so. How it works tells the same client-review story as every other page. The shared button demo and the receiver are defined where they are first mentioned, and the docs open the receiver section with what it installs and the one command that installs it.
Improvement One story, front to back: a change, a client, an accepted result.
Home, Demo, Plans, and Docs now tell one journey: describe the change, involve the person who gets the final say, let an agent prepare it, decide on the exact version, deliver and accept. A real completed run, recorded on 17 September, can be inspected on Verify before starting your own. The invoice examples, the rules rehearsal, and the two-agent discussion stay one disclosure away on Demo; Docs opens with four paths.
0.22.0Feature See the change. Share the decision. Keep the feedback.
Try a visible contact-button change in an isolated ScopeBlind-owned repository. Compare checked file versions, invite someone to review, approve the exact change and accept the receiver-observed result. Requested revisions carry the original signed feedback into a fresh task. Connect your own repository through guided local discovery and a pinned receiver workflow. Scoped agents can inspect and suggest revisions; approval stays with each person. Preview and revision evidence verifies locally in the browser or with verifier 0.10.16.
0.21.0Feature Keep the original intent. Review the exact change.
Start with a short practice, then continue into shared work or a real repository review. Original instructions stay with a signed standard; unresolved meaning needs an explicit decision. Browser author keys use secure non-extractable storage with verified migration. Repository tasks bind both approvals to exact commits, paths and checks; the owner’s receiver keeps the write credential and checks the result at GitHub. Evidence separates integrity, authority, observed effect and recipient acceptance. The gateway and coordination client now share a pinned npm release, while verifier 0.10.15 checks complete nested inputs and distinguishes verified properties from declarations.
0.20.0Feature Start with your agent. Take the next decision on your phone.
A reusable agent profile can prepare an unsigned task and return a private review link. Each person reviews and signs their own limits; negotiation and later execution require separate explicit grants. The original organizer can authorize the same agent for the exact jointly adopted task without replacing its negotiation connection. Authorize a phone key for one task, with limited decision permissions, expiry and revocation. An authenticated inbox shows current work, and optional generic reminders carry no task details or approval rights. Evidence preserves actual device signatures and their original human authorization.
Start through your agentYour decisionsSetup and device scope
0.19.0Feature Bring your agent, resolve a disagreement, compare real plans
Choose Claude Code, Codex CLI or a generic MCP client, with scoped setup and resume instructions. Readiness is confirmed only after the adapter verifies the current signed instructions. Two people can now set hard ranges and private preferences for both budget and review threshold, compare actual gate outcomes, and select one exact tested plan for joint approval. Recovery explains conflicts, offers changes only to your own draft, and links fresh discussions to preserved history or an exact eligible blocked request. No earlier approval or agent authority carries into the new discussion.
0.18.0Improvement Try it alone, return to your work, and share an exact result
Practice with an explicitly scripted partner using real isolated gate comparisons. The shared job now explains each next action and shows the consequences of proposed rules. This browser can restore unsigned drafts and invitations, and a recent-work shelf returns to the last view. People can explicitly share an immutable 30-day negotiation or result snapshot; recipients verify its signed record without an account or approval rights. Private agent instructions and raw invitation secrets stay outside those public copies.
0.17.0Feature Your agent and their agent, with both people in control
Two people can now sign separate mandates, give their assistants private background, and negotiate an invoice review threshold. Hosted assistants or individually paired MCP agents recommend at most three proposals. Every comparison retains both people’s required checks and exercises the actual gate in isolated ledgers. Both humans approve the exact proposal, report, and new agreement before creating a separate trial; the partner is enrolled as its reviewer. The shared negotiation history travels with the final evidence, while private briefs stay outside the shared record. Model work starts explicitly, is bounded, and can recover saved progress.
Improvement One clear path from shared work to a checked result
The public pages now use the shared trial’s visual language and a consistent Try together entry. Demo separates guided work, live-agent jobs, and rules rehearsal. Verify leads with authentic signed sandbox examples. Docs organizes setup by task; Write distinguishes draft, test, sign, and deploy. Plans separates the immediate free sandbox from manual team setup. Data-path explanations now distinguish hosted collaboration, model assistance, local gateways, and local verification. Existing tools, records, and deep links remain available.
0.16.0Feature Test and improve the rules together
Invite a collaborator to test an invoice task without payment authority. Rehearse concrete cases through the database-backed gate in isolated sample ledgers, compare an approval-threshold change, and preserve successful work alongside required safety checks. The owner alone can use a passing comparison to create a separate trial with the proposed rules; the original job remains intact. Dedicated test-agent connections expose the same bounded collaboration through MCP. Optional model readback proposes a case for review and marks unsupported requests. Signed reports bind the source inputs, cases, proposed change, and gate runtime, and verify locally in the browser.
Feature A live agent, your limits, and a shared result
The invoice trial now offers a live model that inspects invoices and purchase orders, proposes payments through the gate, and continues independent work while exceptions wait for a reviewer. Describe the job, review the model’s proposed limits and assumptions, and customize the sample amounts. Requested revisions retain the earlier result, payments, and shared budget. Connect your own MCP agent with a short-lived pairing code, wait for decisions, and revoke its access from the room. All payments remain fictional. The guided walkthrough is still available.
Feature Try a shared job with someone else
The shared invoice trial gives two browsers their own signing keys and a private reviewer invitation. Routine payments run within a shared budget; a larger payment waits for an exact approval, changed details require a new decision, and a duplicate invoice is refused. The recipient can accept the finalized result and export independently verifiable evidence. Requests are guided and scripted; decisions and payments run through a persistent sandbox gate and ledger. No live model or bank is connected. The installed coordination adapter is available as a source-build preview.
0.14.0Feature The record lands on the standard's page; held actions wait there for a named person
A signed standard can be published to a page of its own from the Write page. A gate started with the standard and the page refuses anything over the standard's limit before it runs, holds anything above the approval threshold, and posts every receipt to the page after it is chained locally. The named person approves or denies each held action on the page with the key the standard accepts, signed in the browser; the gate finds the decision when the agent retries the same action and carries it in the receipt. Each run opens on the verify page from the standard's page. The author's key now stays in the browser that made it. protect-mcp 0.14.0 adds --standard, --report, --report-token, and --run to the MCP gateway and the hook server, and passes the call's input to Cedar so a compiled amount limit is evaluated at the gate.
Launch One site at scopeblind.com; the standard as a plain-English document; the writer talks to a model
scopeblind.com and the earlier product site become one site under one name. The standard is rendered as a document a person would sign, from its structure and the compiled clauses, with a mark on every line saying where it is checked; the same document opens above the receipts on a shared record. The plain-English writer now proposes changes through a model (OpenAI or Anthropic, chosen by which key is set) and falls back to fixed sentence shapes offline. Docs, status, and the changelog moved over from the retired console.
Feature Write a standard in plain English, run under it, share the record
A conversation on the left and the standard on the right. Each sentence becomes operations on a structured standard, the compiler turns it into the policy the gate enforces, and the refusals are computed from that policy, never from the words. Sign it in the browser with a key generated there. A checked record can be kept under an unguessable link for 180 days so the other side can check it again without an account.
Launch Stage two: six agents, one bounded allocation, three configurations
Five attested swarm runs published in ScopeBlind/verified-runs. With logs and allowlists alone, nine unauthorized effects reached the services; with a conserved allocation, receivers, and rules over history, zero, at the same cost. Every file behind the numbers is public and checks offline.
v0.10.12Fix @veritasacta/verify 0.10.12 and protect-mcp 0.13.1
One chain-hash rule across every implementation (three conformant: protect-mcp, nobulex, APS), the policy under test is the on-disk bytes, and protect-mcp loads its Cedar engine correctly on Node 18 and 20 after a release that had denied everything there. verify-cli checks a standard against receipts and calls with --standard and --receipts.
Feature Verified runs: a benchmark run every tool call of which is on the record
Run clauses in the compiler, a run manifest, a harness with Codex and Claude hook adapters, and a public repository of runs made in GitHub Actions under the maintainer's published key, with the model answering inside a confidential machine that signs every call.
Feature The standard compiles to the gate's policy
A signed standard compiles to protect-mcp's Cedar policy with a digest every receipt cites; real gateway receipts verify in the browser on the verify page and with one CLI.
Feature Proof Requests and admission decisions
A recipient states the decision, the boundary, the evidence that would change it, and what happens if it is met, signed in the browser. An operator's payment is evaluated against it before dispatch and after, and the admission decision is recomputed from the evidence by anyone.