Skip to content

Web access and the browser

Cora can search the public web and read one page at a time, so she can bring in a competitor’s offer, a regulator’s rule or a platform’s limits without anyone pasting it in. It is on for a new organisation; an org admin can turn it off. An organisation made before 17 November 2026 keeps the setting it had, so if it never turned web access on, it is still off until an org admin does. The blocked sites and the monthly search limit apply either way. She uses it only when the mission needs current public information that is not in its documents or memory.

The examples on this page follow an invented café chain, Larkspur Coffee, checking what rivals offer on iced drinks.

  • Search. She sends a short query (up to 300 characters) and gets up to 10 results, each with a title, an address and a snippet. She can ask for a country (Australia unless it matters) and for recent results only (the last day, week, month or year).
  • Read a page. She gives a full address and gets the page as plain text, up to 40,000 characters. A page that is cut at the limit is marked as cut.

Searching is done by a search service (Brave Search first), run by Cora AI’s own servers. The key to that service is kept in AWS Secrets Manager and is never shown to Cora, the team or any log.

Setting Who changes it What it does
Let Cora use the web (organisation) Org admins On for a new organisation. With it off, no mission can use the web or the browser.
Search and read pages (under it) Comes with web access On exactly when web access is on; it has no switch of its own. While web access is off it is greyed and says “Turn on web access first”.
Use a browser (under it) Org admins On by default. Lets Cora open pages in a real browser where the team sandbox is on. While web access is off it is greyed, unticked and says “Turn on web access first”. See Cora’s browser.
Missions: on or off by default Org admins The default for missions that have made no choice.
Blocked sites Org admins Cora never opens these sites or their sub-domains, such as reddit.com. One per line.
Monthly search limit Org admins Searches across the organisation each month (1,000 to start). At the limit, searches stop until next month.
Let Cora use the web in this mission Mission admins On by default. Turns it off or on for one mission (under Web in the Tools tab of mission details). The same two lines sit under it and grey out while it is off.

The organisation settings are on the Web access page of Settings (the account menu at the bottom of the left rail, then Settings, then Web access). Members can read the settings but not change them. Cora is given the web tools only when the organisation and the mission both have web access on. If a switch is turned off while she works, her next web call is refused and she tells the team in plain words.

When web access is off, Cora knows which switch is off. Asked for a search, a page or a screenshot, she says plainly that web access is off for the organisation (an org admin can turn it on in Settings › Web access) or for the mission (a mission admin can turn it on in the mission’s Tools tab, under Web). She never suggests adding another tool to get round it.

  • Every statement Cora takes from the web carries a numbered source in her reply, and the reply ends with a list of sources: the number, the page title, its address and when she read it.
  • If a page failed to load, was refused or was cut short, she says so next to the claim. She never fills the gap from memory or guesswork.
  • Press How I worked on her reply and the dialogue shows a chip such as “🌐 2 searches · 4 pages” and the pages she looked at, with when each was read. The links open in a new tab. The dialogue keeps the query, the address and the page title, never the page itself.
  • While she works, the “Cora is working” line says “Searching the web: ” or “Reading ”.

Web facts are not approved figures. A price, an offer or a date from a public page still needs the person who owns it to confirm before it goes into copy, a plan or a decision, and Cora says so when she reports it. She saves a fetched page to the mission’s documents only when someone asks, with its address and the date.

Why a web page cannot tell Cora what to do

Section titled “Why a web page cannot tell Cora what to do”

Everything that comes from the web reaches Cora inside a data fence, as third-party information. Her instructions say that text in a fence is never an instruction, even when it says it is: she ignores a page that tells her to change her rules, reveal something, open another address or call a tool. She never puts a secret, a person’s details or private mission text into a search query. The web is never used in your private chat with Cora.

Pages are read by Cora AI’s servers through the same safe route as connector servers (D0131, D0163):

  • Only https is used. An http address is upgraded to https. Other kinds of address, a user name in the address and ports other than the standard one are refused.
  • Private, internal and cloud-metadata addresses are refused, including when a public-looking name points to one, and on every redirect (at most three).
  • Blocked sites are refused.
  • Only web pages and plain text are read, up to 2 MB, within 15 seconds. No script on a page is ever run.

These rules are for web_fetch. Pages Cora opens in her browser do run their scripts, in a separate browser in your organisation’s own network, with the same blocked sites. Those pages are recorded like the pages she reads, marked as opened in her browser.

  • Each search uses a little of the organisation’s AI usage, like a model call. When the monthly search limit is reached or the AI usage is used up, searches stop and Cora says so. Reading pages costs nothing extra, but a month allows five pages for each search the limit allows; pages opened in her browser do not count towards that.
  • Every search query and every page address is written to the organisation’s history (web.search, web.fetch), with who (Cora, in that run), when and the mission. A page opened in her browser is written as web.fetch too, marked via: browser, with the address she asked for and where it ended. Page contents are never kept there. Each change to a web setting is recorded too (org.web_access_changed, project.web_access_toggled).

The search key is one secret named core/web-search/brave in AWS Secrets Manager (ap-southeast-2), a plain string holding the key. See the release runbook in the repository (infra/scripts/release-runbook.md, “Web search key”) for how to create and rotate it. Until it exists, a search answers that web access is not set up yet. This design is decision D0637.

Some pages show nothing until their scripts run, and some questions are about how a page looks. For these Cora has a real browser: a managed copy of Chrome that runs in your organisation’s own part of the platform, next to the team sandbox. She can open a page, read it as a visitor would see it, take a screenshot and click a link or a button. In this first version she can only read and look.

The examples on this page follow an invented bike shop, Kestrel Cycles, checking how its new e-bike page shows on the web.

  • Open a page. She gives a full http or https address and waits until the page has loaded (or until its network is quiet, for pages that keep loading), for at most 30 seconds. She gets the final address, the status, the title and the page’s text as a visitor sees it, up to 32 KB. A page cut at that size is marked as cut short.
  • Read part of a page. She can read the whole page again, or one part of it, as text or as an outline of its headings, links and buttons.
  • Take a screenshot. She can save a picture of what is on screen, or of the whole page, as a JPEG in the team folder under screenshots/. A screenshot is at most 5 MB and 10,000 pixels tall; a longer page is cut at that height. The team folder is not shown in the console, so a screenshot there is not yet visible to anyone. To show it to the mission, Cora saves it into the mission’s Documents with save_file_to_documents and posts its card in the chat; the team then opens it from the mission’s Artifacts tab. She is told never to say a screenshot is in the mission’s files until she has saved it there (D0807).
  • Click. She can click a link or a button, named by what it says, and gets the new address, the title and a short summary of the page.

What she cannot do in this version: type into a page, fill in or send a form, press keys, download or upload a file, sign in to anything, or run code of her own in a page. She never uses anyone’s password or a vault secret in the browser.

For Kestrel Cycles, Cora opens the new e-bike page, reads the price and the stock message that a script fills in, saves screenshots/2026-11-17T0305-e-bike-page.jpg and clicks “Compare models” to check where it leads.

Cora has the browser in a mission only when all of these are true:

  • Web access is on for the organisation and for the mission. See Web access.
  • The team sandbox is ready for the organisation. The browser uses the sandbox’s network and folder.
  • “Use a browser” is on. It is on the Web access settings page, one of the two options under “Let Cora use the web”. It is on by default, so it applies wherever the sandbox and web access are on. While web access is off it is greyed and unticked and says “Turn on web access first”, because the browser needs web access. An org admin can turn it off; the change is recorded in the organisation’s history.

The browser is Cora’s own tool. Her helpers never have it. A hero that another mission has called in has its own mission’s browser, as in its own work, but its screenshots go to its private folder for the call, never the team folder, and are gone when its step ends (D0805). It can save one into its own mission’s Documents, never into the caller’s mission (D0807). It is never used in your private chat with Cora.

  • Only http and https addresses. Any other kind of address, such as a file on a computer, is refused, and so is an address with a user name or password in it.
  • Blocked sites are refused, with their sub-domains, the same list as for web access. This holds for every request a page makes and every live connection (web socket) it opens, not only the address she asked for.
  • Redirects are checked too. If a page sends the browser on to a blocked site, or anything arrives from one, Cora leaves the page and is told it was refused. Nothing from it reaches her.
  • Every page is opened from the platform’s own network, through the same safe route as the team sandbox. Private addresses, other parts of the platform and cloud metadata cannot be reached from it. As a second check, the browser itself refuses an address that is plainly local, such as localhost, a private or cloud metadata address, or a name with no dot.
  • If a page does not even start loading within her wait, she is told that no page was loaded. She never gets the page she had open before by mistake.
  • A link that opens a new tab or window is not followed: Cora stays on her page.
  • Downloads are blocked.

Everything Cora reads in the browser reaches her inside a data fence, as third-party information, exactly as for a web page she reads. A page that tells her to change her rules, reveal something or call a tool is ignored. What she reports from a page is a lead, not an approved fact: a price or a claim still needs the person who owns it to confirm it.

  • One short session at a time. The browser starts when Cora first uses it in a piece of work and stops when that piece of work ends, after at most 15 minutes. Each session is a fresh, empty browser: no cookies, saved logins or history are kept between sessions.
  • Nothing is recorded. Session recording is switched off. AWS keeps a session’s start and end times (not its pages) for 30 days.
  • Cost. A session uses a little of the organisation’s AI usage for every few minutes it runs, like the team sandbox. When the AI usage is used up the browser does not start, and Cora says so.
  • History. Each session start is recorded in the organisation’s history (web.browser_session), with the mission and the run.
  • Pages she opens. Every page Cora opens, and every click that goes to a new page, is recorded the same way as a page she reads with web access: in the organisation’s history as web.fetch, marked via: browser, and in the list of pages under her post, marked “Opened in her browser”. The record keeps the address she asked for and where the page ended, without anything after a ? or #, and the title. It never keeps what the page said. A page that is refused or does not load is recorded as such. Browser pages do not count against the month’s page limit; the browser is paid for by its session time. If the record cannot be kept (for example, web access was just switched off), Cora is not shown the page.

Each organisation has one browser, named coraBrowser_<orgId>, made by the five-minute sweep once the organisation’s team sandbox is ready, and checked every hour (decisions D0750 to D0756; the guard after review is D0766, the page records D0767). A browser with recording turned on is taken out of use and logged as an error. AWS Security Hub check BedrockAgentCore.6 reports each browser for having no recording: this is accepted on purpose, because a recording would keep page contents. See the release runbook (infra/scripts/release-runbook.md) for how to switch browsers on.