What changed
What's new in MeshArc: 13–19 November 2026
Better-ranked search with DuckDuckGo alongside Google and Bing, word counts when you compare any two runs, more dependable screenshots, and MCP server 0.7.10.
Search
Results ranked for what you asked. When a query names a version, results about that version now lead: "Postgres 18 release notes" no longer opens with the Postgres 15 notes. One site can no longer fill the top results; its third result onwards moves a few places down, unless you named that site or limited the search to it. Nothing is added or removed, only the order changes.
DuckDuckGo joins Google and Bing. A search asks Google first, then DuckDuckGo, then Bing, and merges their answers by relevance.
A region when you name none. A search that names no country or language now gets US results from Bing and worldwide results from DuckDuckGo, instead of results for wherever our server is. A search that names a country or language is unchanged.
Timings in every answer. Each search reports how long it took (tookMs) and how long each engine took (ms on every attempt), so a slow search shows where the time went.
If you watch a search with a monitor: because of the new ranking and region, its next run may report some results as moved, new or dropped. After that one run it reports only real changes.
Comparing runs
Word counts for any two runs you pick. Comparing a run with any earlier one now shows each changed page's words added and removed. When a page is not counted, because the comparison reached its limit or the text is no longer kept, it says "words not counted" instead of showing +0 and −0.
Runs read differently. When the two runs were read under different extraction settings, or after an update to how MeshArc reads pages, the comparison says so in plain words. A page whose words are the same in both is counted as unchanged rather than listed as changed.
Screenshots
A capture that hangs gives up and tries again. A screenshot now gives up after 10 seconds and tries once more, instead of waiting 30 seconds and failing. When no picture comes, the page says why in plain words.
Fairer pricing when no picture comes. If a screenshot was the only reason a page was rendered in a browser, on a site that doesn't need one, and no picture came back, the page is charged as an ordinary fetch.
Projects and pages
Archived projects are marked as archived and can be restored.
A site's own server error reads as one, not as a page we failed to read.
Captcha pages are recognised again when a reCAPTCHA or hCaptcha challenge carries its own hidden answer field.
MCP server and SDKs
MCP server 0.7.10, with 23 tools, including screenshot_url for a picture of any page and archive_project. The hosted server at mcp.mesharc.dev runs it already.
search_webcan search news (source="news"), with each result's publisher and age, and a dropped connection is retried without searching or charging twice.get_changesexplains a comparison between two runs you pick, including pages whose words were not counted.screenshot_urlreports the credits the capture actually used.Clearer input checks and fixes from our connector testing.
SDKs. The Python client is at 0.7.10, with screenshot(run_id); the Node client is at 0.7.2.
What's new in MeshArc: October and November 2026
Web search, monitors and an AI agent join the crawler; sign-in with Google or GitHub; scanned PDFs and Word files read as text; and a round of reliability and security work.
Web search, monitors and the AI agent
Web search. Search the web and get ranked results, with each page as markdown in the same call if you ask for it. Keep a search to the sites you name, or search news with the publisher and age of each result. A search costs 4 credits for its results page; the same search again within the hour is free.
Monitors. Keep a search on a schedule. Each run is compared with the last, so you are told what is new, with an optional monthly credit cap.
AI agent. Ask a question and the agent searches, reads the pages and answers with its sources, as text or as JSON in a shape you give. Each run stops at the credit cap you set.
More dependable search. A results page that is refused is retried on paid plans, and a search that found nothing says so and is not charged.
Crawling and content
Scanned PDFs and Word files. A PDF with no text layer is read by OCR, and Word (.docx) files are read as headings, paragraphs, lists and tables.
Documents behind a bot wall. A PDF on a site that challenges plain requests is now fetched with the cookies a browser earns there, at the usual document price.
Cleaner main content. One-off scrapes and crawls drop sidebars, related-post blocks, link menus and footers that a site built without marking them. Lists that belong to the content, such as a record's tags, are kept.
Clearer answers when a page is not read. A site that does not exist, a page a plan limit stopped, and a short page that needed a browser each say what happened and why.
Sign-in and accounts
Sign in with Google or GitHub, for new and existing accounts.
Email preferences. Choose which emails you get under Profile; every optional email has a one-click way to stop it.
Sub-accounts on Scale and Enterprise: workspaces for your own customers, created and managed through the API.
New workspaces can no longer be opened with a disposable email address.
Webhooks, SDKs and MCP
Webhook subscriptions. An app can register its own webhook beside a project's, and a search can have a webhook too. A project's webhook can carry each changed page's markdown.
SDKs. The Python client is at 0.7.3 and the Node client at 0.7.0, with search, monitors and the agent. A LangChain package is available.
MCP server. One-click install buttons for the common assistants. A search or a small scrape now returns its result in the same call, a repeated scrape reads the page again after ten minutes, and counts come back as numbers.
Plans and credits
Stealth trial on the free plan. Free workspaces can try the residential and unlocked tiers on 200 credits.
Limit and credit notices. When a plan limit stops a page, the API and an email say so. A free workspace is told when its credits run low and when they run out.
Credit caps per run and per month, for projects and monitors.
Reliability and security
A run carries on across a deploy, and the same request sent twice starts the work once.
Lists and reads are faster, and requests no longer wait behind slow scrapes.
Fetches, redirects, webhooks, destinations and your own model endpoint reach only public addresses.
An API key limited to some projects is held to them on every project route, including delete.
Dependencies were updated for security fixes and locked to the versions production runs.
What changed, 1 October 2026
Admin
A blog manager role, and the operator's screens behind their own gate.
AI assistants and search engines
Llms.txt, the docs as text, an IndexNow key.
Auth
Decide where a sign-in lands by resolving it, not by its prefix.
Where a sign-in lands has to be a path on this site.
Blog
A post names the ones out before and after it.
Import posts from a spreadsheet.
Act on many posts at once, and a schedule never unpublishes.
A post cannot take an address the blog's own pages use.
Posts, the image library, scheduling and the public API.
A blog manager role, and the cleaner every post goes through.
Authors in the sitemap, and each list dated by its newest post.
Previous and next under each post, and related posts called that.
The import screen.
A content security policy for the blog and changelog, report-only.
The newest posts on the home page, and a page's own on each tool.
The link dialog finds our own posts and pages.
The posts list filters, pages and acts on many at once.
The blog manager's screens -- posts, the editor, images, calendar.
The public blog at /blog.
Changelog
Drafted from each deploy, public once a person approves it.
The public page at /changelog.
The approval screens.
Consent
A stale factor, and one person in two workspaces.
A lapsed session signs in again, and a new account comes back.
A page that connects an app, and copy that no longer says "local".
The banner reads the cookies without setting state in an effect.
Ask where the law says to, and remember where a visit came from.
Copy
The MCP server offers seventeen tools, not sixteen.
Crawl
A wall answering the not-found probe is not learned as the not-found page.
Destinations
A one-off crawl or scrape writes where it names.
Legal
The policy says what the site actually does.
Newsletter
The list, with double opt-in.
The subscribers screen.
The sign-up, and the pages its emails open.
OAuth
An idle TTL of zero or less keeps the month; the doc says what revoke ends.
A key appears when the app connects, and the log says what happened.
Revoking the key in Settings ends the connection, and the deploy works.
One approval, one ordinary API key.
OAuth and scrape follow-ups
Polling reports the destination, keys record use.
Playground
An extraction may name a destination too.
Send the result to a destination, not only to the screen.
Proxy
A file upload passes through as bytes, and the API skips the page gate.
Rate limits
Count the address Cloudflare saw, not the one a client claims.
safeNext
Refuse a path that dot segments turn into another host.
Share cards
Every tool, use case and comparison has its own.
Sign-up
An account remembers where it came from.
Other
Max tier auto, the plan as the ceiling everywhere, and climbs bounded by the credits left.
Sign-in keeps an app's consent link, and CI type-checks with route types.
Max tier auto in the app, its docs, and billing screens that say what the plan allows.
Fix(web): send /signed-out to /login on the public host, not localhost.