Releases

Big audits publish again, and two false positives are gone. Reports now stay under the publish limit no matter how many pages you crawl, you get warned before starting a cloud audit you can't afford, and a failed publish refunds the whole audit. Plus a new check: soft 404s.

Added

• New check: soft 404s. Pages that return HTTP 200 but actually render "page not found" content (framework error shells, not-found titles) now get flagged. Search engines treat these as thin or duplicate content, and they usually mean a route is silently broken. Detection is conservative: it requires multiple strong signals before flagging, so a short page with an unlucky title won't trip it.
• Content and legal checks skip error pages. Rules like cookie consent and content quality no longer judge a page that is really a 404 in disguise. That was the source of phantom "missing GDPR consent" warnings on sites whose error pages leaked into the crawl. Skipped pages are shown as skipped, not silently dropped.
• Know the cost before you spend. Cloud audits now estimate their credit cost up front (flat base plus per page) and warn you, or ask for confirmation in a terminal, when your balance won't cover it, instead of burning credits on a run that can't finish the way you wanted.

Fixed

• Large reports publish again. Publishing a big audit (hundreds of pages) could fail with "report exceeds 5MB" even though the real limit is 20MB. Two bugs, both fixed: the error message quoted a stale limit, and site-wide findings embedded every crawled URL, so payload size grew with crawl size. Findings are now sampled (up to 100 pages per finding, with a count of how many more there are), which keeps any report, from 10 pages to thousands, comfortably under the limit. The CLI also checks the size before uploading and trims further instead of failing after the crawl.
• Cloudflare Turnstile is no longer invisible. Forms protected by Turnstile (and other CAPTCHAs) using the explicit-render or preloaded loader pattern were flagged as unprotected. The check now recognizes preload hints and empty widget mount points, so protected forms pass.
• Failed publishes refund the audit. If your report can't be published because of a size limit or a server error, the entire audit's credits are refunded automatically instead of leaving you charged for a report you never got.

Credits stop being a mystery: every audit now shows exactly what it cost and why, and your org gets a full transaction log. Also the first release to arrive on Windows through the repaired self-updater: if you're on 0.0.74, this update installs itself.

Added

• See exactly what every audit cost. Cloud audit reports now include a cost breakdown: the flat audit base, each rendered page, cache hits, and any refunds. It appears on the report page in the dashboard, and agents get it too: get_report over MCP includes a cost summary, and a new list_credit_transactions tool pages through your org's full credit ledger. There's also a new Transactions view in dashboard settings showing every grant, debit, and refund. If your per-audit cost seems to swing, this is where you see why: only pages whose source changed get re-rendered, so audits right after a deploy render (and bill) more pages than audits of a quiet site.
• squirrel feedback learned categories. An interactive picker (or the --category flag) files your note as a bug report, feature request, praise, confusion, missing data, or tool ergonomics, and records which CLI version it came from. The same categories run through the CLI, MCP, dashboard, and website, so feedback lands in one place no matter where you send it.
• Delete websites for real. A new delete_website MCP tool (with dashboard support) removes a site and immediately frees the slot against your website cap.
• Dashboard polish. Tracked issues can be filtered by how they were found (full crawl vs smart audit), and page-limit settings now show the effective cap when a plan ceiling clamps a custom value.

Fixed

• squirrel self doctor no longer cries corruption over a permission error. Running doctor from a directory it can't read (locked-down CI images, restricted shells) misreported healthy settings as corrupt and exited non-zero. Permission problems on the working directory are now reported as exactly that, and the rest of the checkup runs normally.
• MCP sessions no longer drop when tokens refresh concurrently. Two clients refreshing the same session at the same time could race and knock each other out; a short reuse window now lets both land on the rotated token.

Faster

• Repeat cloud audits skip redundant summary work. Summaries for pages unchanged since the last run are served from a content-keyed cache instead of being recomputed, so re-audits finish sooner.

Pick your rules and skim your results: audits gain include/exclude filters and a summary mode, local dev crawls get dramatically faster, and Windows self-updates stop dying mid-download. Plus a batch of sharp-edge fixes across auth warnings, settings handling, and cloud publishing.

Added

• Filter which rules run with --rule-include / --rule-exclude. squirrel audit --rule-include ax,performance runs only those categories; --rule-exclude images skips them. Bare names cover a whole category, category/rule targets a single rule. The report is marked partial and the health score recomputes from what ran. Typos, contradictory filters, and --fail-on gates against excluded categories all error before crawling instead of wasting a run.
• --summary for a quick score check. Console output trims to the score, category breakdown, and issue counts with no per-issue detail: handy for CI logs and fast re-checks. Works on squirrel audit and squirrel report.
• Localhost audits are much faster. Plain-HTTP loopback targets (your dev server) automatically skip the polite per-host crawl delay, and new --concurrency / --per-host flags control parallel fetches. Public sites keep the respectful defaults.

Fixed

• Windows self-updates no longer die mid-download. The background updater ran as a detached child that Windows kills the moment the parent command exits (Job Object semantics), so updates silently never landed. The update now runs in-process with a bounded grace period at exit, and downloads are atomic: a killed update just retries next run.
• A corrupt or unreadable session file now warns loudly in every command. Cloud-touching commands explain the session could not be loaded and how to check it, instead of silently running anonymous. When an environment token authenticates the run anyway, the warning is suppressed entirely: no more false "running anonymous" on authenticated CI runs.
• Settings lookup no longer mistakes permission errors for missing files. An unreadable parent directory now surfaces a clear error instead of silently skipping your local settings.
• Windows users get pointed at the right installer. Running install.sh from Git Bash or MSYS now prints the PowerShell one-liner instead of a dead end.
• Cloud publish reliability: oversized robots.txt content is clamped instead of rejecting the whole report, and a timing window where a just-published report's run could still be reaped and refunded is closed.

Hosted MCP grows up: sessions renew themselves instead of expiring, connecting shows exactly what you're granting, and agents get a direct feedback channel. Plus a friendlier installer and a stack of cloud reliability fixes.

Added

• Hosted MCP sessions no longer expire out from under you. Connecting over OAuth now issues rotating refresh tokens, so a session renews itself silently instead of going stale and demanding a browser re-consent every 30 days. Any replay of an already-used refresh token revokes the whole session family on the spot.
• The MCP consent screen shows exactly what you're granting. Approving a client now lists each permission (run audits, view credits, create API keys) with individual toggles, and the required read access is clearly marked. A client that asks for nothing specific no longer receives a full-access grant by default.
• Agents can send feedback mid-session. A new send_feedback MCP tool takes a category and a message, optionally tied to a run or website, so an agent can report confusing output or missing data the moment it hits it. Works with read-only credentials.
• The terminal report now leads with the big picture. squirrel audit output opens with the four top-level group scores as a bar breakdown, and categories are ordered most-severe-first, matching the other report formats.
• A friendlier install. The install script greets you with the CLI's own banner and finishes with a numbered get-started guide: first audit, agent skills, cloud login, plus shell completion and squirrel self doctor hints.
• Failed installs can now tell us why. The install scripts report a failure's platform, step, and a sanitized error line (paths scrubbed, nothing personal) so broken installs get fixed fast. Reporting never blocks or slows an install, and setting NO_TELEMETRY disables it, same as the CLI.

Fixed

• A cloud audit that wedges mid-run is now failed at its own deadline and refunded. Stuck running audits were previously reaped by a blanket 75-minute backstop; each run is now judged against its own configured budget, fails promptly with an honest message, and refunds automatically whichever path notices first.
• Quick audits stay quick. The post-crawl phase (site metadata, technology detection) now respects the audit's overall time budget instead of adding up to four unbounded minutes to a fast crawl.
• Rendered audits no longer lose cookies. A header-handling bug made cloud rendering drop every cookie a site set, breaking pages behind cookie-dependent front-ends and blinding the cookie security rules. All Set-Cookie headers now survive the full render path.
• CLI settings writes are now atomic. A crash or full disk mid-write can no longer corrupt settings.json; settings are written with owner-only permissions, and a permission problem reading credentials is reported loudly instead of being treated as logged out.
• Crawls cut short now say so. A crawl interrupted by the time backstop gets a distinct "stopped" status: its collected pages are still analyzable, and squirrel report explains the crawl stopped before finishing instead of leaving an ambiguous state.
• Max pages per audit now follows your plan on every path. A couple of API paths could start an audit above the plan's page ceiling or estimate below it; dispatch and pricing now clamp consistently, and a stored setting above your plan's ceiling is honored at the ceiling.
• Publishing a report with an extremely long list of findings on a single check no longer fails the audit; oversized lists are trimmed safely with a note, keeping the rest of the report intact.
• Toggling a website's badge on and off in quick succession now always lands on the state you chose, even when the clicks race.
• When an audit fails because the CLI that started it disconnected, the failure message now says exactly that instead of a generic infrastructure error.

Repeat audits get dramatically faster and cheaper: unchanged pages now skip cloud rendering entirely, and rendered pages come back as each one finishes instead of waiting on a whole batch.

Fixed

• Repeat audits reuse renders the way they were always meant to. Two bugs kept the unchanged-page check from ever engaging: sites that send no cache validators never recorded a content fingerprint, and sites whose HTML embeds a millisecond timestamp produced a different fingerprint on every fetch. Both are fixed, so a re-run of an unchanged site now skips rendering, and its per-page render charges, outright. In testing, a repeat audit's crawl phase dropped from minutes to about a second.
• Rendered pages stream back as each one finishes. Render results were previously held until an entire batch completed, so one slow page delayed every page behind it. Fresh crawls of render-heavy sites are noticeably faster, helped by quicker render queue pickup and more rendering headroom per browser.
• A page with an extremely long identifier can no longer fail the whole report. Publishing rejected a report when a check item's identifier (often a very long data: image URL) exceeded an internal length cap, failing the audit. The CLI and the server now shorten these safely while keeping distinct items distinct.
• Cloud audits interrupted by infrastructure hiccups now retry themselves. A cloud audit whose container never started or died mid-run previously failed on the spot. It now gets one automatic redispatch before giving up, with guards that stop a superseded container from ever overwriting the retried run's results. A retry that also fails still refunds automatically.
• Brief platform blips, like a dropped database connection or a rate limiter restarting during a deploy, are retried once instead of surfacing as errors in the dashboard, the API, and the MCP server.
• The dashboard no longer crashes for users browsing with Google Translate or similar extensions that rewrite the page.

Agent experience goes deep: a much larger set of rules auditing how AI agents read, reach, and act on your site, plus major crawler and cloud reliability fixes.

Added

• The Agent Experience category grows from 4 to 17 rules. New checks cover agent access (whether GPTBot, Claude-User and friends get the same content as a browser, including bot-challenge and pay-per-crawl detection), content signals and licensing (contradictory bot policies across scopes, noai signals, RSL licenses), agent-facing files (AGENTS.md, llms.txt including lookalike SPA shells, MCP server cards, A2A agent cards and other well-known agent endpoints), API discoverability (OpenAPI, OAuth self-onboarding), and response token weight for agents on a budget. The crawler now probes these agent surfaces on every audit.
• New security rules: subresource integrity on external scripts, and cookie security flags checked against real response headers. The catalog now totals 261 rules across 21 categories.
• Report issues are now ordered by severity in every output format, so the most important findings always come first.
• Failures outside your site's control, like an unreachable third-party link, are now reported as warnings with an expected-failure tag instead of counting against the audit as errors.

Fixed

• Cloud audits that stalled mid-crawl and timed out now complete normally. A crawler concurrency slot could leak when a page fetch failed a certain way, eventually deadlocking the crawl until the run hit its time limit. This was the main cause of recent cloud audit timeouts; affected runs were refunded automatically.
• Crawls stop promptly and cleanly at the page cap instead of letting in-flight work run past it.
• Publishing a report with the full rule catalog could be rejected after the catalog grew past an internal limit. The limit is raised and now guarded by tests so it cannot silently recur.
• Auto and hybrid render modes no longer render pages that were already rendered during the crawl, making rendered audits cheaper.
• Reports now say when a cloud check did not run and why, instead of leaving a silent gap in the results.
• Set-Cookie headers now flow through the whole audit pipeline, so cookie security rules evaluate the real headers your site sends.
• Creating a new site through a cloud audit or the MCP server now respects your account's website limit like every other path.

A dashboard and reliability release: clearer audit status, honest failure reporting, and safer invites.

Fixed

• Your dashboard issue count and Issues page now agree. A website card could show a large issue count while the Issues page showed none, because the two read from different places and the issue list could lag behind the latest audit. The Issues page now tells you when a sync is pending and gives you a one-click resync from your most recent report.
• Website thumbnails no longer hang on "Capturing…". A screenshot that failed to capture would spin forever and read as a broken empty box on every visit. Captures now report their outcome, so a finished thumbnail appears right away and a failed one falls back to a clean placeholder instead of an endless spinner.
• Blocked audits read honestly in the dashboard. When a site's bot protection or firewall blocks the crawler, the report now explains that the site blocked the scan and how to let it through, instead of an unexplained grid of zeros or a misleading "Starting" label.
• Signed-in audits are tracked reliably and respect your site limit. An audit started while signed in that failed to register no longer runs silently untracked, and creating a brand-new site through an audit now honors your account's website limit like every other path.
• Audits that fail to start now report the failure and refund any credits reserved for them, closing gaps where an early failure could go unrecorded.
• Repeat audits no longer risk matching a page against a stale content fingerprint, keeping cached-page reuse accurate across runs.
• Organization invite requests made with an API key are now authorized against that key's own organization and scope, not the account that created the key.

A reliability hotfix: audits that stalled without producing a report now run normally.

Fixed

• Audits that hung and collected no pages now work. After some upgrades, a project's local database could be left without a column the crawler writes on every page, so each page silently failed to save. The audit then crawled until it hit its time limit without ever producing a report, spending render credits along the way. squirrel now repairs the missing column automatically the next time it opens the project, so affected sites audit normally again. If you saw an audit run for many minutes and then fail with "no pages collected", this is the fix.
• Audits now stop quickly when every page fails to save or fetch, instead of grinding all the way to the crawl time limit. A systematic failure now fails fast with a clear message rather than burning time and credits.

A reliability release: accurate reports on large crawls, honest failure reporting, and faster repeat audits.

Added

• New --fresh-ua flag on audit and crawl to re-roll the browser identity the crawler uses for a project.

Changed

• The crawler now picks its random browser identity once per project and reuses it on every run. Repeat audits hit the render and analysis caches far more often, making re-runs faster and cheaper.
• JSON reports now include status and statusReason, so a failed or blocked audit no longer reads as a clean pass to scripts and agents consuming JSON output.

Fixed

• Large crawls keep every affected page in the report. On audits of hundreds of pages, a rule that flagged many of them could have its list of affected pages silently cut off past an internal limit, dropping pages from the published report and skewing the score. Those pages now fold into one accurate finding, so big audits report and score every page they should.
• Quick-coverage audits on paid plans no longer blame a cloud outage for checks that quick mode intentionally skips. The report explains it was a quick scan, shows how to run the full set, and renders the Agents score as locked instead of silently missing.
• Sites behind Cloudflare and similar bot protection that answer with a 503 challenge page are now reported as blocked with actionable advice, instead of a generic "site unreachable".
• Audits started with a bare domain (squirrel audit example.com) now reliably appear in your dashboard run history. They ran fine before but could be invisible to run tracking.
• Debug logs print real error details instead of [object Object].
• The crawl command docs now describe the --coverage flag and the real default page budget.

Simpler pricing and faster repeat audits.

Changed

• New flat pricing: 50 credits per audit plus 2 credits per rendered page. Everything else that runs inside an audit is now included: AI content analysis, authority signals, technology detection, the editor summary, site metadata, domain stats, ad-block detection, dead-link checking, and report publishing. Keyword and content gap analyses stay optional add-ons at 25 credits each. A 50-page rendered audit costs exactly 150 credits, and the estimate you confirm up front is the price you pay.
• Publishing is always free. Publishing a report at any visibility, including flipping an existing report to public later, never costs credits.
• Failed audits are refunded automatically. If an audit fails or is cancelled before delivering a report, its charges are returned, including the base.
• Audits need a balance of at least 50 credits to start. Below that, the CLI runs the audit locally, tells you why, and skips cloud features instead of charging you partway.

Fixed

• The end-of-audit credits line now matches your ledger. It is built from the amounts the server actually charged, including the audit base, instead of client-side estimates that could overstate spend.
• Crawler user agents are modern browsers again. The random user-agent pool no longer includes decade-old browser versions that bot protection loves to challenge, so crawls of protected sites are far less likely to stall or get served challenge pages.
• Lower memory use on large audits. The audit engine now releases parsed pages it is not actively using, cutting the working set on 100+ page runs and avoiding severe slowdowns on memory-pressured machines.

New

• Per-phase timing breakdown. Run with --debug to see exactly where audit time goes (crawl, cloud analysis, rules, report). The same breakdown is stored with the run, which makes slow-audit reports much easier to diagnose.
• Repeat audits are much faster. Unchanged pages reuse cached rendering and AI analysis server-side. In our testing, re-auditing an unchanged 100-page site dropped from about 40 minutes to about 4.

A reliability release: honest reports when a site blocks the crawler, cloud audits back at full power, and you now hear about it when an audit fails.

Fixed

• Cloud audits ran without rendering and cloud checks. Audits triggered from the dashboard were silently locked out of browser rendering, technology detection, and the cloud checks. On sites with bot protection this often meant an empty, failed report. Dashboard audits now run with full capabilities.
• Blocked sites now say so. When bot protection, a firewall, an auth wall, or rate limiting refuses the crawler, the report is now marked blocked and explains what happened, with concrete next steps: allowlist the crawler, turn off the blocking rule for the audit, or run the CLI from a trusted network. Previously these audits published an empty report that could read like a clean pass. The agent-facing report formats carry the same signal, so your coding agent knows the audit was blocked instead of concluding your site has no issues.
• Failed audits no longer show "No issues found." A report with zero crawled pages renders its failure state instead of a success summary.

New

• Audit failure notifications. When an audit fails, you get a dashboard notification with the reason and a link to the affected website. Failures are also tracked on our side so we can spot problems before you report them.

Changed

• robots.txt is no longer enforced by default. Audits are run by site owners, so Disallow rules and Crawl-delay no longer apply to your own audit. robots.txt is still fetched and parsed for sitemap discovery and the robots audit rule. Opt back in with respect_robots = true in your config; when enforced, Crawl-delay is capped at 2 seconds so a slow directive can't stretch an audit into minutes of waiting.

The Team plan arrives, plus clearer auth guidance from the local MCP server.

New

• Team plan. Invite teammates into an org at $29/seat/month (2-seat minimum, billed monthly), with a shared pool of 3,000 credits per seat. Assign roles (admin, editor, viewer, billing) and manage seats from Settings → Team in the dashboard.

Improvements

• Clearer auth errors from the local MCP server. Calling an authenticated tool without a session now returns an actionable error naming both fixes: set SQUIRRELSCAN_API_KEY for headless and CI use, or run squirrel auth login for an interactive session.

squirrelscan v0.0.64

• Four new performance and image rules (249 total): LCP images without fetchpriority="high", slow web font delivery (font-display, preload), hidden carousel slides eagerly loading images, and rendered vs intrinsic image aspect-ratio mismatches.
• Four score groups: every report now rolls categories up into SEO, Performance, Security, and Agents scores, with a redesigned report UI and shareable OG cards to match. Category scores also weigh how many items each finding affects, so one big issue no longer counts the same as one tiny one.
• Fewer false positives across the rule set: service-area businesses without a street address pass LocalBusiness checks, phone matching understands country codes and trunk zeros in tel: links, FormShield counts as form protection, license banners no longer flag minified JS, keyword-stuffing ignores stopwords and repeated CTA buttons, sitemap coverage respects the crawl page cap, and hidden carousel slides are no longer told to lazy-load above the fold.
• Hosted MCP server: point any MCP client at https://mcp.squirrelscan.com and your agent gets the full toolset with OAuth, no local install. Tool responses got smarter for fixing agents: rule lookups now include a concrete recommendation and a docs link, and audit status reports live progress and a completion reason.
• Free cloud audits got the full treatment: logged-in free accounts now run the same scan pipeline as Pro, differing only in included credits.
• New tutorials: step-by-step guides for fixing your site with an AI agent, running local audits with the CLI, and going deeper with cloud audits at https://docs.squirrelscan.com/guides.

See https://docs.squirrelscan.com for documentation and the full rule set.

squirrelscan v0.0.63

• New install endpoint: curl -fsSL https://install.squirrelscan.com | bash now installs from a dedicated, edge-cached delivery that no longer rides on the marketing site. The old install URL keeps working.
• WAF challenge detection: audits flag DataDome and Kasada challenge pages that answer with a 200 instead of a real block, so a bot wall no longer looks like a healthy page.
• Bigger cloud audits finish: raised the cloud audit time budget so larger sites (~20+ pages) complete instead of timing out.
• Better domain data for country-code TLDs: WHOIS lookups fall back to a secondary source when a ccTLD registry (e.g. .au) times out.
• Clearer reports: pages that were render-blocked and recovered via a direct fetch are now called out in the report.
• Faster, fairer per-host crawling, plus internal hardening and reliability improvements.

See https://docs.squirrelscan.com for documentation and the full rule set.

squirrelscan v0.0.62

• Audits of down, blocked, or empty sites now report an honest N/A score and a failed/blocked status instead of a misleading "A / 100%" — across console, JSON, Markdown, LLM, text, and XML output.
• Internal audit-engine hardening and reliability improvements.

See https://docs.squirrelscan.com for documentation and the full rule set.