What's New

The Cactus mark rests on a hand-drawn ink line with grass wherever it appears, from the site header and footer to Messenger launchers and share cards. Its foot is squared beneath the line, while raster favicons and reply-card tiles keep the round foot.

Building on Live setup progress, each build card shows an expected finish time based on repository size, set at kickoff and refined twice as the build runs.

A ready page opens from the catalog without a reload. If its link isn't in the current list, Cactus checks the catalog again before saying the page doesn't exist.

Building on product setup finishes after pages publish, Cactus shows your site's brand-color choices as light-and-dark pairs in Product branding and saves both colors when you confirm. A dark color appears only when it has enough contrast against Cactus's dark background.

Dark mode gives the main card its own fill and glow above the near-black sidebar, with a shadow on wider screens.

Building on live setup progress, a single line at the bottom of the sidebar shows how long your build has left on every page, then offers a way home when it finishes.

Building on Help Center pages in more languages, you can add or remove available language packs under Settings › Languages. Saving the list asks Cactus to check translations, and conversation translation is on by default.

On larger screens, the home page pairs its hero message with a still Cactus button that opens Messenger. On non-touch screens, the Messenger launcher waits until the hero scrolls out of view before appearing.

Building on A visual guide to How Cactus Works, the home page puts the primer and links to useful settings in front of you while your product's first build runs. The primer explains why Cactus's view of your product is grounded only once the build is running.

During setup, choose your Help Center and Intercom conversations together in one step. Your Intercom choice stays selected as you continue to connect it.

The planted Cactus mark follows Kevin’s drawing down to the ink line, with a shorter ground and no pebble. Building on A light-only Cactus website, the artwork appears in the logo, launcher, hero, footer, app icon, and share card.

You can clone a public Help Center as an Agent Memory Repo, extending Markdown pages for AI agents with its pages, topics, and changelog in one repository.

Building on Product-grounded Messenger answers, Cactus remembers a person's earlier conversations and uses relevant ones to answer their next question. You can turn memory off in Messenger settings. The team owner can review or forget what Cactus remembers in the Inbox.

Turn translation on under Settings › Languages. Inbox conversations appear in your team's language, with the original text folded into a quote above the translation until you open it. Cactus answers in the customer's language and translates messages both ways.

Help Center pages, topic names, and Messenger text are translated into each language your team adds, alongside English. Corrections keep their answer beside the first, while screen labels and terms stay as written.

After a customer proves who they are, they can share, rate, or hand off their own thread from a second device.

Building on Cactus checks code when facts fall short, Cactus checks proposed answers against Help Center pages and saved knowledge before writing them.

When a visitor asks from a help screen, Messenger captures that screen before switching to chat. A failed question's Retry resends the screen context and the pictures attached to the original question.

In the Help Center, pages, What's New, and releases scroll within one full-height card while the site's bar and Cactus's column stay in place. A page opens at its first line, while lists return to where you left them.

Each team's Inbox shows turns only for its own products, keeping other teams' customer conversations private.

A handoff from the Help Center, Messenger, or an AI client sends the team a link to the conversation in the Inbox and thanks the visitor. Products with a connected Discord destination also offer Talk to a human.

Subheadings separate ideas, numbered steps lay out procedures, and tables and notes organize details. A rewrite can leave out sections that don't help the reader.

Help Center pages give AI agents clear links, readable state labels, and correctly nested headings in their machine-readable text. An Agents page at /agents explains how to connect an AI agent to your Help Center, and you can turn it off in Connect under Appearance > Help site.

When a visitor chooses “Talk to a human” on your Help Center, their conversation appears alongside Messenger and AI client threads in the Inbox, where your team can reply.

On a first build, each page's name and first 12 questions guide a search for matching files, which join the evidence the writer already has. Files too large to include whole can contribute excerpts instead of being left out.

On a full build, Cactus maps each screen, integration, settings group and public policy by product area, tracking what still needs evidence. A code push refreshes the index, and another push during that refresh gets one follow-up pass.

Each new commit becomes a part of the open daily release, and Cactus plans and writes the release when the day closes. The feed, release pages, Help Center cards, Messenger and Slack show releases as text rather than cover images.

Attach an image to a question in the app or Messenger, and Cactus can consider it when answering. You can also add an image to a Rules entry or a help page section.

The Visibility card puts search results in plain language: how many phrases were checked, whether your sites appeared in the top 10, and which watched products ranked instead. A table lists each phrase and the best positions for your sites and watched competitors.

The Visibility dashboard charts mentions, citations, model trends, competitors, and traffic to your sites. You can resolve or dismiss recommendations and read summaries of what changed and what to do. Checks query one AI model every other morning.

What's New groups releases by theme, with headline chapters and a rollup of smaller changes. The same structure appears in timeline answers from Cactus and Slack release cards.

The AI clients page lists clients and agents signed in as you, with Disconnect controls. Browserless agents can be claimed using a code in Cactus, and removing a teammate blocks their AI-client access on the next request. Cactus also exposes agent tools and A2A on its site and Help Centers.

An unsettled question reaches Slack with evidence, and you can answer from the card or Truth. Cactus uses that answer on the next pass, and corrections can propose writing rules for approval. Teammates see who set cited rules, while owners control changes to other members' rules and pages.

Cactus checks each promised question against its answer section, keeps fulfilled promises, and raises gaps for thin answers. Single-page rewrites and push passes preserve those gaps. Visitors get a page only when it answers their question fully. Otherwise, Cactus can combine up to three pages.

The Help tab shows the site's catalogue with a filter. Select a result to open its full page, including sections, figures, related links, and previous and next pages.

Choose a client and select Open in Claude Code to give your coding agent a command containing your personal token. Run the command, then restart the client if it asks. The page masks your token and lets you revoke it.

The app, Help Center, and Messenger use the same release cards and monthly feed. Open a release to expand its details where it appears, and follow links to the Help Center pages it changed.

Agents can request Markdown versions of Help Center pages and find resources such as llms.txt, the API catalog, and authentication guidance.

At the brand step or in Settings, tell Cactus who reads your help pages. In Settings, you can also choose a figure style. Cactus uses the audience you set when writing each page.

Help pages can include evidence-based diagrams, screen sketches, rails, comparisons, and highlighted terms. Cactus chooses where a figure helps, and answers can show a page’s figures in the app, Help Center, and Messenger.

Point a custom hostname to Cactus with a CNAME record, then follow its verification and certificate status in Connect. Cactus serves the Help Center from that address.

The AI clients page gives you a user-level install command and hands sign-in back to you. Cactus offers each caller only the MCP tools their access permits, and outside clients receive readable answers with links to cited help pages.

In an Inbox conversation, add an internal note, ask Cactus, or draft a customer reply in the composer. A draft stays unsent until you choose to send it, and replies render as formatted prose in Messenger.

Your first code search runs its queries even if an AI client sends an unknown page cursor. Cactus asks you to search again only when a cursor it issued is no longer available.

Use the Published and Source switch above a Help Center page to inspect its live sections, supporting code, unanswered questions and applied Rules. Source view is available to your team in the app, not to public Help Center visitors.

Connect Slack from onboarding or Settings under Handoffs. In a handoff thread, teammate messages become private notes unless you ask Cactus to reply to the customer. Cactus follows threads after answering and can investigate product code or refresh a help page there.

Connect a Discord server from onboarding or Settings under Handoffs, then use /cactus ask, /cactus status, and /cactus link. Linked teammates get answers with their team's access, while visitors get visitor answers and linking instructions. Long answers include the sign-in line within Discord's message limit.

Settings gives you a team MCP address, one-click links and configuration snippets for supported AI clients, plus guides for Claude and ChatGPT. A separate address lets people ask about a product's published Help Center without signing in. Teammates can approve, deny, or revoke access for connected clients.

From a conversation thread, you can ask Cactus to save its answer as a help page for you, your team, or everyone. Cactus sends the request through its regular page-writing process, using the audience you chose. Pages meant for you or your team are written for that audience instead of being withheld as unpublished.

In Settings, “Learn from my writing” creates rules from your product site and imported help and conversations. Each rule includes the sentence it came from, so you can review the guidance alongside rules your team wrote.

Set a page's audience to only you, your team, or everyone. Keep the Help site team-only while you review it, and copy a public link from any page visitors can see. Pages set to only you stay out of teammates' views.

When a projection changes a page's title, summary, sections, or sources, Cactus saves a dated edition with its reason. Open an earlier edition read-only or use Show changes to compare its wording with the live page.

Intercom imports up to 1,000 of your newest closed conversations from the past year, leaving out bot and workflow replies. If an import fails, its status explains what happened and offers a retry. Setup also identifies app review as the reason a connection can't proceed.

From Installations, open Claude Code, Cursor, or VS Code with the install prompt ready, or copy the prompt for another coding agent. The prompt gives the agent framework-specific install steps and a checklist for verifying the launcher.

Your public Help Center has a What's New page with monthly stories and releases, linked from the site header. Visitors can open each release from the timeline.

Cactus turns help-site topics into open questions, and brings past conversations and replies in as dated evidence. When an import surfaces unanswered questions, Cactus can draft pages for those gaps and check them against the pages you already have.

Build history and Rebuild are in Settings, where you can review each run's stages and published-page count. Repository pushes start one build per day, with later pushes folded into a scheduled run.

Before a batch rewrites pages, you can preview it, cancel it, or start generation. Rule changes also update pages for the chosen audience, and Settings reflects the result of applying or retiring a rule.

Choose whether people join by domain policy, invitation, or owner approval. Owners can invite any email address, manage domains, and rename the teamspace. Failed invite emails appear on the member row.

When you hide a concept, Cactus withholds it from external surfaces. A shared answer page that cites it is also hidden from visitors, while your team can still see the thread.

Questions the Help Center can't answer can trigger a code investigation. The answer lists the files it drew on for your team, while visitors get a single customer-safe source line.

Cactus answers for the product linked to the channel, with the complete answer and links to cited Help Center pages. Follow-up mentions in the thread continue the same conversation.

Attach a web page as a product source, and Cactus can use it alongside code when answering questions. It checks the page when it answers from it, and marks related help sections for refresh when the page's text changes.

Messenger shows the steps in a reply as Cactus works, then streams the answer in readable text. If the turn fails, customers see the failure and can retry; when it finishes, the answer appears with its source cards.

A visitor's turns stay in one conversation, which appears in the Inbox with their email if they leave one. When they ask for a person, it moves to Waiting on you and the team gets an email they can reply to. Teammate replies and customer feedback stay with the thread.

Choose Combine into one page at the end of a thread with two or more answer pages. Cactus folds the questions and answers into one page, labels how many questions it covers, and shows that page when you share the thread.

Each product has its own Messenger key, saved design, and list of allowed website addresses. Cactus keeps the settings with the product and checks the key and site when Messenger loads or handles a request, so a release won't wipe the setup.

On pages with an inline conversation, Cactus opens in the page without an overlay, and the answer streams there while chat stays conversational. If you leave, the conversation moves to the reading column and returns when you come back. On other pages, Cactus opens beside your cursor.

When the evidence doesn't support a step, Cactus leaves it out rather than adding a disclaimer. For deeper topics, it can connect related concepts in an operational walkthrough, with citations tied to lines in the files it read.

When a new build refreshes a page, sections from the previous edition can be kept, rewritten, merged, or removed to fit the new writing brief.

Questions from a Help Center page stay in an inline chat: your questions and Cactus's live thinking sit on the left while the answer takes shape on the right. Follow-ups stay in the thread, and answer pages are clickable cards.

Copy a public link to an answer from the inline chat. Use Save as a page when you want the answer without its conversation history.

Cactus reads your teamspace's company domain for branding and keeps the accent color you choose. The completion card waits until help pages publish, and connecting GitHub takes you back to the product you left.

With a work email, Cactus creates a teamspace named after you or offers a one-click join when your email domain already has one. Naming it for you keeps the teamspace distinct from your product's brand.

When you connect GitHub, Cactus verifies installations you control through GitHub's authorization return, without asking you to link your GitHub identity to your Cactus account. If you haven't installed the GitHub App, the flow takes you to its install page.

Each conversation shows its status and labels. From there, you can follow it, escalate it, or resolve it.

Question threads show pending and error states. If a live answer doesn’t come through, the thread can still display its saved reply, and you can retry when generation fails.

Each product gets a What's New timeline alongside Docs. It records product releases after each build so customers can look back at what changed.

The setup chip opens onboarding with a repository picker. Your product home lists published help pages.

Customers see a question box beside a list of concepts, and Cactus renders each answer as a page with its sources.

When a published Help Center page covers a question, Cactus answers from its text and cites that page instead of searching product code.

Each Help Center page has a machine-readable Markdown version, so AI agents can read the same published guidance.

Cactus generates customer-safe Help Center pages from product knowledge after each knowledge update, organizing them by category and concept with citations. AI agents can use the guidance without reading product code. When citations become stale, you can regenerate one concept at a time.

The Product Map connects customer questions to code concepts, then leads Cactus to definitions and references in your product code. Code also gets more weight than prose when search ranks evidence.

Choose which packages in a monorepo Cactus maps, then change that selection from the Product Map.

Cactus adds investigation findings as a new entry in the Product Map, leaving existing entries intact. If an update fails or is interrupted, the map keeps its existing state.

Start in Concepts to see the entities most of the map depends on, with their connections described in plain language. Switch to Graph to explore relationships visually, or select an entity to inspect its summary, code evidence at the pinned commit, and connections grouped by verb.

The cursor draws as soon as its position is known, before you move the pointer. Presence chips match cursor colors, the navigation rail gets a slimmer sliding marker and anchored popovers, and Messenger Studio has segmented controls, a concise install page, and an Alderly sample scene.

Controls and panels use solid fills, with a fine inset line for depth. The text cursor takes a brighter coral color so it stays legible against dark surfaces.

A self-contained visual guide shows how Cactus works, alongside a plain-language account of its behaviors, numbers, rules, and sharp edges. Color distinguishes when Cactus acts alone, when a person is needed, infrastructure and facts, and production measurements.

The changelog drops its cards and restores the glow on its status light. Its status can finish updating instead of staying stuck on “building,” and longer updates can complete without an answer-generation error.

In Conversations, open and escalated badges use distinct fills. Disabled reply actions are clearer, text has stronger contrast across the main screens, and answer lists scroll without overlapping.

Cactus saves facts from its research, so repeat questions can reuse what it learned instead of repeating a full code investigation. When your product's first build is underway, a knowledge gap can get an honest in-progress reply.

Cactus marks core knowledge ready after researching the product and its primary customer outcome, then continues deeper research in the same build. If a documentation crawl fails, the build can continue with sources it did capture while the docs source is deferred.

Eligible Help Center pages publish automatically when a build produces customer-safe prose. You can still unpublish a page, and Cactus keeps that choice across later builds until you explicitly publish it again.

In Knowledge, browse a page tree, switch between Customer, Team, and Dev views, and inspect the answers behind each section. Recent additions link to the page they changed, and Settings lets you turn automatic gap-filling on or off.

Cactus groups automatic changes from Git, Jira, and Linear into one evidence-backed product update every 12 hours. Manual refreshes, onboarding, and facts you teach still update product knowledge immediately.

During setup, Cactus separates core product understanding from full readiness, so you can see when the first grounded knowledge is ready while the rest of the build continues.

When a product answer is incomplete, Cactus can route the request to a person through your configured Handoffs. The chat keeps showing investigation progress as Cactus works toward an answer.

Cactus can use sections of its product knowledge alongside individual facts as evidence. That lets it synthesize an answer even when no single fact covers the whole topic.

When someone asks for a person, Cactus routes the request to your team even if the message also includes a product question or contact details. After a teammate replies, Cactus stops answering until the conversation is handed back.

Messenger shows the answer as Cactus writes it, so you can start reading before the full reply is ready.

Each refresh ties product facts to the source snapshot they came from. Cactus keeps an unaccepted refresh out of customer-facing answers until its validation checks pass.

For customer-facing answers, Cactus uses facts that pass checks for evidence, freshness, relevance, and audience visibility. If those facts don't support an answer, Cactus withholds it instead of trying again with broader context.

Cactus matches links and action routes to evidence from the product’s published pages. When that evidence changes, it refreshes affected answers so those routes stay aligned.

With a published Help Center, open Docs or What's New from Messenger’s action bar. Read pages and release entries in place, then ask Cactus about an article with it still in context.

While Cactus works through a question, Messenger shows phase-by-phase activity and distinguishes a deeper code investigation with its own card. Collapse the steps into a compact summary when you’re done following along.

Cactus matches the customer's question against saved resolutions and serves a close match. If customer-ready product knowledge is missing or fails to load, Messenger hands the question to your team instead of guessing.

Messenger uses one message identity for streamed answers and the saved conversation feed, so an answer appears once instead of as a duplicate. Its live feed also restores messages when a customer closes and reopens Messenger.

In the Inbox, you can reply to customers, add private notes, ask Cactus to research or draft a reply, and resolve or reopen a conversation. Teammates can see who else is viewing it, while status updates carry through to Messenger and Slack.

When a guest escalates without an email on file, Messenger can collect their address in the conversation. The Inbox keeps the contact with the thread so the team can follow up.

A teammate can reply to a customer's conversation from its Slack escalation thread, and the answer returns to Messenger. Slack's request list links back to conversations, while email links can reopen the matching conversation in Messenger.

The Inbox uses a three-pane workspace for the conversation list, transcript, and customer details. Message receipts show delivery and reading status for the latest exchange, and unread conversations stand out in the list.

When a visitor shares an email, Messenger attaches it to the anonymous conversation they already started. Their email or name appears in the Inbox conversation list and transcript.

Open All Answers to search the catalog of answers, see whether a question was customer-asked or generated, and check when it was updated. Suggested questions come from your product's concepts and open in the cursor's ask flow. For a faster first build, onboarding can use a shorter changelog window.

Each answered question has a permanent link to its live public version. Copy the link from the answer reader to share it.

In the answer reader, links to your product's own pages open within Cactus. Links to external sites open a new tab.

Public release updates use names and concepts from your product's own vocabulary, keeping internal codenames out of what customers read.

Cactus checks whether its evidence directly answers the question, then searches relevant code when facts are thin or only adjacent. Your product map, Rules, and lessons shape replies, which stream in chat with inline citations.

Customer-facing answers translate code-derived behavior into plain language and keep source paths, internal routes, architecture terms, and private names out of the prose. When code shows how a feature works, Cactus can still explain that behavior without exposing implementation details.

When Cactus starts a deeper investigation in chat, you can follow its current step and see the result when it finishes. The investigation's answer appears in the conversation instead of stopping at an in-progress status.

In Connect, choose a Messenger launcher face and sound pack, then preview them alongside the conversation. Messenger loads your saved sound pack when it starts.

Select a source chip to open its cited facts and scores. Screen readers announce streamed replies, and open and close sounds work in both Messenger modes. A Powered by Cactus signature appears when that branding is enabled.

Setup progress appears as soon as a refresh is queued and shows indexing, mapping, building, and ready with live counts.

Connect feature-flag sources in Sources to include them in the product knowledge Cactus uses.

The publish review groups new, changed, and removed answers so you can approve the set before customers see it. Open an answer's history to read earlier versions, or annotate it to correct a fact or guide how it appears.

The preview keeps launcher presets available in cursor mode and gives the launcher a pressable keycap. It uses the embedded Messenger's own cursor, so the preview shows a single cursor.

Cactus uses relevant, product-facing facts to answer and cites the concepts behind its response. Investigation failures and unverifiable claims stay out of its product knowledge.

Setup gets past connector and source-index blockers, keeps audience selection consistent, and displays readable chat bubbles. Missing or unavailable organization logos appear as fallback avatars.

Use Reset in onboarding or from the catalog page to clear setup and generated product knowledge. Your connected sources stay in place, so you can rebuild from those same repositories and connections.

A three-pane reader pairs your product's concept tree with claims and their code locations, confidence, and history. Its review queue surfaces conflicts, low-confidence claims, and unanswered questions, which you can address by fixing the code or adding authoritative guidance.

When a fact used in an answer changes, Cactus refreshes that answer against current product knowledge. It can detect distinctive values such as prices, limits, URLs, and error messages even when an answer leaves them uncited. Fresh answers are reused for repeat questions, including common contractions.

When a customer asks for a person in Messenger, Cactus records the handoff in the Inbox and sends it through your configured notification channels, such as Slack or email. If delivery fails, the request stays recorded in the Inbox for your team.

Compatible MCP clients such as Cursor and Claude Desktop can search Cactus product knowledge, find similar answers, inspect concepts, and ask product questions. You can use these tools from internal MCP clients, too.

Cactus can create or reuse #cactus-escalations when you connect Slack, making it the product's handoff destination. You can also create it from the channel picker, and existing connections can reconnect to grant the required channel permission.

The index matches code and documentation by meaning and exact terms. It filters out files made up only of credentials and redacts inline private keys before indexing, while keeping code searchable without summarizing every file.

Cactus stores each source summary with a citation to at least one real passage. When generated citations don't match indexed content, it falls back to the source's first passage.

Cactus plans repository scope before refreshing the source index, so indexing follows the repositories selected for the product.

Generated explanations use customer-facing names for connected sources and leave internal source labels, repository-indexing terms, and drafting artifacts out.

Answers describing your product draw on audience and customer-workflow facts, while leaving distribution and route details out.

Timeline updates draw on refreshed product facts and use customer-facing language. They avoid unsupported timing and universal claims, and leave code-like evidence details out.

Cactus checks source evidence before treating a product detail as current. It then brings product knowledge into line as evidence from connected sources converges.

Generated help pages organize product information around reader tasks and show concise context about their sources. They keep setup details and unsupported claims out of public pages.

Atlas builds each page from connected knowledge entries instead of exposing raw implementation and debugging terms. Dynamic answer pages also no longer repeat the title and answer in the body.

Choose Preview in Portal from the layer switcher to see public Atlas documentation and the Timeline changelog with your product's branding. Atlas carries your progress into the next session.

Cactus connects product entities to repository evidence and plans ingestion around repository scope and tests as it builds product knowledge.

The guided flow moves from repository selection to the Linear connection, then to your product’s brand details. If you leave to authorize access, it brings you back to finish onboarding.

Cactus adds relevant Linear context to shipped-change narratives. When connected work details change, Cactus refreshes the affected narratives.

The install guide gives you framework-specific instructions and code for HTML, Next.js, React, and SvelteKit. It shows where each snippet belongs, fills in your widget key when one is available, and includes checks to confirm installation.

On Atlas, the in-app guide walks you through connecting GitHub and choosing repositories, then presents a brand card for you to confirm.

Atlas groups shipped changes into chronological clusters, each shown as a card. You can switch timeline layers, and a loading indicator appears while the data loads.

Choose Developer, Internal, or Customer in the layer switcher. The assistant follows that choice when it searches or answers, and a brief indicator shows when you switch.

Dock the in-app assistant in the bottom-right corner, or drag it out when you want it back on the page. Double-click its cursor, choose the dock button, or type /dock to switch. Escape can re-dock it after a conversation.

Related changes are grouped by area and time along one continuous timeline. Each card includes a quick overview, a structured walkthrough, and individual changes to inspect.

When you choose a monorepo in Connect, Cactus reads its workspace configuration to find packages and select the ones to analyze. Package analysis and build history stay scoped to each package’s files.

Add Rover to your Discord server from Connect, then link the server to your Cactus organization. Direct mentions get a response, while Rover can skip follow-ups in subscribed Slack or Discord threads when they don't need its input.

Rover's Slack replies use richer formatting, with feedback buttons, an Atlas link, and product context. Slack's App Home shows your product overview, recent changes, and customer Inbox. @mentions still reach Rover when Slack delivers overlapping events.

Slack replies can call tools for up to five steps and show progress as Rover works. Memory tools can retain user context across conversations.

The MCP server includes a full what-shipped lookup, product-change retrieval, and investigation tools. Compatible clients can filter shipped changes by date and receive progress notifications while tools run.

Switch Timeline between Dev, Internal, and Public to see change stories for each layer. Those stories draw on recorded product changes and can be translated between layers.

Customers see refined cursor and brand interactions in the widget. Messenger conversations appear in the team's Inbox alongside conversations from other customer channels.

A valid authorization header lets Cactus open a replacement session after an expiry or server restart. MCP clients can use search and ask, inline citations display without a blocked label, and conversation history loads through the threads endpoint.

A keyword match can bring in related code even when it doesn't match a function or class. Search uses real file content and keeps same-named files from separate repositories distinct.

MCP uses one Streamable HTTP endpoint. API keys control access: dev keys have full access, internal keys exclude code, and public keys can be shared and tracked.

The MCP search and ask tools accept a repository ID, so you can limit a query to one repository. Results also follow the access layer on your API key.

The catalog has a tab for each layer, with a status indicator. In the internal layer, search code by question, and ask the chat widget about the page you're viewing using its interactive elements as context.

The catalog marks pages as pending, draft, or verified in their cards and page view. An area shows draft if any page inside is a draft, and shows verified only when all its pages are verified.

Repository selection and setup status appear as interactive blocks alongside the conversation. Your chat history and choices persist when you return.

The catalog playground groups sample areas and pages into a browsable view, with code references beside each page.

The response arrives in a steady flow, making the back-and-forth in chat feel more fluid.

Use the new sign-in flow to access Cactus.

The What Cactus does page has a two-line subheading and a refreshed design with a dot grid.

You can join the Cactus waitlist.

Product setup includes an early GitHub integration.

The setup flow adds buttons and categories, and shows progress as you work through it.

The branding pairs indigo-purple with orange and uses updated fonts.

A wider design pass polishes the product and Help Center’s overall appearance.