Hyper docs

Goal for this release

The last release brought computer use, Windows support, and more ways to work across connected tools.

This release makes that work faster, easier to navigate, and more interactive. It brings new models, Hypercards, five new connectors, a redesigned desktop interface, and a rich-text editor for automation instructions.

Our goals were to:

Make opening the app, switching conversations, searching, and reading replies faster.
Let you edit and send emails through an interface built for the task.
Turn tables and media into richer results inside the conversation.
Make connected accounts and their capabilities easier to understand.
Make automation instructions easier to write and maintain.
Preserve your input and completed work when a task stops or reaches its limits.
Improve connector reliability and make failures easier to diagnose.

New models and a faster Hyper

We upgraded:

GPT-6 Sol → GPT-6.1 Sol
Claude Sonnet 5 → Claude Sonnet 5.5

GPT-6.1 Sol is now the default for new chats. The model picker keeps its existing order, and you can still choose another model.

Saved Sonnet 5 selections resolve to Sonnet 5.5. (#887, #937, #972)

Recent conversations reopen faster

Switching back to a recent conversation no longer always requires downloading its entire transcript again.

Hyper keeps a small cache of recently opened conversations and invalidates entries when their contents change. The cache is cleared when the live event connection ends, rather than continuing to show copies it cannot keep current.

The first opening of an uncached conversation still loads its history from the server. (#918, #927)

Faster knowledge search

Knowledge search now uses an index to find matching content without scanning every document.

For common terms, Hyper first checks recent files and uses that result only when older files cannot change the returned answer. Less common searches fall back to the full indexed query.

This improves search speed without changing account access, source filters, or the ordering of results. (#957, #967, #978)

Smoother streamed replies

Replies now appear a word at a time with steadier pacing.

Formatting arrives with the text it belongs to, instead of briefly exposing unfinished Markdown such as bold markers or partial links. Tables and other incomplete structures wait until there is enough content to display them correctly.

New words fade in, with reduced-motion settings respected. (#1068)

Less work before the first screen

We reduced startup work by loading optional session-replay code only when it is needed and avoiding an up-front scan of the operating system’s certificate store.

Code highlighting also loads individual language grammars on demand instead of loading every grammar before displaying the first code block.

Loading animations do less work on the main thread and pause while the window is hidden. (#912, #916, #928, #931, #938)

Introducing Hypercards

Hypercards bring rich, interactive interfaces directly into your conversation.

Instead of treating every result as a block of text, Hyper can give you an email to edit, a table to explore, or media to view inline.

Draft, edit, and send emails in chat

Email compose cards support Gmail, Outlook, and Resend, on desktop and iPhone.

You can:

Edit To, Cc, and Bcc recipients.
Change the subject.
Write and format the message.
Review the attached files.
Send the email from the card.

For example:

Draft a follow-up to this customer and attach the proposal.

Make this email shorter and change the subject.

The card keeps the draft in the conversation while you review it. (#1060, #1072, #1075, #1081, Hypercards plan)

Drafts stay editable

Edits save to Hyper and remain available when you reopen the conversation. Other devices can load the updated card, and Hyper receives your changes when you continue the conversation.

You can also ask Hyper to revise a draft rather than creating another unrelated email.

The compose card shows the sending account and attachments. Changing the sending account or adding and removing attachments directly in the card is not part of this release. (#1082, #1084, #1085, #1086, #1091, #1094)

Sending stays explicit

An email presented as a compose card does not send until you click Send.

This also applies when you have an Allow always grant or YOLO mode enabled. Those settings determine whether another approval is needed after clicking Send; they do not skip the compose card.

If Hyper cannot confirm whether an email was sent, the card tells you to check before sending again rather than offering an automatic retry.

Automation runs keep their existing action behavior; they do not stop at these compose cards. (#1083, #1087, #1093)

Richer tables and media

Tables in replies can appear as cards with controls to sort, filter, and copy their contents.

Hypercards also bring images, videos, and maps into the conversation. Supported standalone links can become embedded views instead of remaining plain links.

These are presentation features, not permission to change the underlying records in another service. (#1090, Hypercards plan)

Five new connectors

Hyper now connects to Trello, Firebase, BigQuery, Google Analytics, and Google Ads.

The four Google connectors previewed in the last release have received approval and are enabled for stable builds.

Supported operations remain limited by the connected account’s permissions, provider restrictions, and Hyper’s action controls. (#941, #1055)

Trello

Connect Trello to search and manage boards, lists, and cards.

Hyper indexes open boards and their open cards, including card descriptions, board and list names, labels, due dates, and links. Webhooks and periodic synchronization keep supported changes current.

Comments, checklists, attachments, members, and custom fields can be read on demand rather than copied into the background index.

Supported actions include creating and moving cards, updating lists, adding comments and checklists, and managing board membership. (#941, #944)

Firebase

Connect Firebase to work with Firestore, Realtime Database, and project settings.

Hyper can read documents, run supported queries, inspect database settings, and perform supported data and project-management actions.

Firebase is read on demand rather than copied into a background index.

This connector does not include Firebase Authentication users, Storage files, Hosting, Functions, Remote Config, or Messaging. (#1017, #1055)

BigQuery

Connect BigQuery to query your data with SQL and work with datasets, tables, schemas, and jobs.

Hyper reads current data when a task needs it rather than indexing warehouse contents. Read-only queries have bounded results and a default limit on bytes scanned.

Supported actions also cover data changes, dataset and table management, and load, copy, and export jobs. Queries can incur Google Cloud charges, and write operations follow their approval requirements. (#1018, #1055)

Google Analytics

Connect Google Analytics to run reports across your Google Analytics 4 properties, including real-time reports.

Hyper can also inspect accounts, properties, data streams, key events, custom dimensions and metrics, and supported settings.

Supported management actions include changes to properties, streams, key events, and selected product links. Universal Analytics is not included.

Google Analytics is read on demand, and newly collected data remains subject to Google’s reporting delays. (#1018, #1055)

Google Ads

Connect Google Ads to review campaigns, ad groups, ads, keywords, budgets, costs, and conversions across accounts you can access.

Hyper can also work with supported campaign, budget, bidding, and conversion-management actions.

Google Ads is read on demand rather than indexed in the background. Changes to live campaigns, bids, and budgets can affect spending immediately, so review those actions carefully. (#1018, #1055)

A redesigned desktop experience

Search and settings have their own space

Search and Settings now open in the main area instead of appearing over the conversation.

The sidebar provides direct navigation between chat, search, and settings. Returning to a conversation preserves its draft, pending attachments, and reading position.

The conversation list is now labeled Chats, with a dedicated button for starting a new chat. Sidebar selection, contrast, and navigation controls have also been refined. (#899, #902, #904, #910, #914, #917)

One question and response at a time

The conversation now focuses on one question and its response.

Use the table of contents to open earlier turns. Scrolling over the navigation rail moves through older or newer turns without requiring a separate click on each one.

This changes how the conversation is displayed, not what Hyper remembers. The complete transcript remains part of the conversation. (#1040, #1062)

Request changes where you are reading

Select part of a reply to open a Request changes field beside the selected passage.

Write your feedback and choose Add to Chat. Hyper adds the request with the relevant passage, so you do not have to copy the text or explain which part you mean.

Selection handling is also more reliable, including paragraph selection and dismissing the feedback control. (#888, #890, #1051)

Clearer tool activity

Tool arguments now appear as labeled chips instead of a raw JSON block.

Open a chip to inspect longer values, nested fields, or highlighted code. Arguments can also appear while Hyper prepares a call.

Provider-supplied readable reasoning summaries remain in the activity timeline after a step finishes and when the conversation is reopened. Script commands and embedded code are easier to read with improved highlighting. (#922, #930, #933, #935, #936, #940)

Automation instructions get a rich-text editor

Automation proposals and saved automation settings now use a formatted editor instead of a plain text box.

You can write:

Headings and paragraphs.
Bulleted, numbered, and task lists.
Bold, italic, and strikethrough text.
Links.
Quotes, inline code, and code blocks.
Dividers.

Instructions remain stored as Markdown. Opening an existing automation does not rewrite its instructions or mark it as changed. (#945, #958, #1037)

Format and rearrange instructions directly

Select text to open the formatting toolbar, or type / to choose a block type.

Block handles let you drag instructions into a different order. The block menu also supports changing the format, duplicating, moving, and deleting a block, with keyboard shortcuts for common actions.

These edits support undo. (#968, #986, #1003, #1010, #1030)

Better pasting and validation

Pasted Markdown preserves supported headings, lists, formatting, and line breaks. Pasted spreadsheet cells remain separated instead of running together.

The editor flags instructions that exceed the limit and prevents submitting empty instructions.

On iPhone, settled automation cards display formatted instructions. Editing automation instructions on iPhone still uses the plain-text field. (#1036, #1037, #1044, #1022)

Better connector settings

Connector settings now show one tile per service, even when several personal or teammate accounts are connected.

You can filter the catalog by category, combine category filters with search, and find services in alphabetical order.

Open a connector to see:

What Hyper can read.
Which supported actions it can perform.
What is indexed and what is read on demand.
Important access limits and risks.
Connected accounts and their sharing settings.

Connect, reconnect, disconnect, and sharing controls live in the same detail panel.

These changes make existing access easier to understand; they do not grant additional permissions simply by opening the panel. (#948, #954, #980, #997, #999, #1006)

Easier access to your work folders

Hyper now keeps a list of allowed folders on each desktop instead of a single work folder.

Add or remove folders in Computer settings. An existing work-folder selection carries over into the list.

When a task needs files in another folder, Hyper can ask directly in the conversation:

Choose folder: Open the system picker and allow the selected folder.
Skip: Continue without granting a folder.

The request can be viewed on another device, but the folder is chosen on the desktop that will run the commands. (#921, #923, #926)

Scratch space for temporary work

Commands without a selected working directory now start in temporary scratch space.

Commands using your files must start inside an allowed folder. Hyper checks the resolved path before starting, including paths that pass through symbolic links.

Allowed folders control where commands can start; they are not an operating-system security sandbox. Desktop scripts still run with your local account’s permissions, and script approval rules still apply unless YOLO mode is enabled. (#919)

More capable YOLO mode

Optional YOLO mode now covers:

Connector actions in interactive conversations.
Connector actions in automations.
Computer scripts in both chats and automations.

While enabled, actions that would otherwise ask for approval can run without another confirmation. This can include sends, updates, deletions, and scripts that change files.

An automation action explicitly set to draft still stays a draft.

YOLO mode does not supply information only you can provide. Folder requests, sign-in and reconnect flows, and questions still wait for you.

Turning YOLO mode off restores normal approval behavior for subsequent actions, including during an ongoing run. (#1019)

Hyper for Intel Macs

Hyper now has a dedicated Intel Mac build alongside the Apple silicon and Windows versions.

The download page offers the appropriate Mac version and lets you choose when the browser cannot identify your processor. Automatic update feeds also distinguish Intel and Apple silicon builds.

Download Hyper at heyhyper.ai/download.

Intel Mac support first shipped in 1.8.2 and is included for anyone upgrading from 1.8.0. (#891, #892, #897, #907)

A rocket game while you wait

Click the rocket above the composer while Hyper works to open a playable game.

A status indicator keeps you informed about the task, including when Hyper finishes or needs your input. Return to the conversation from the indicator, Back to chat, or Escape.

Closing the game does not stop the agent.

Finished games can appear on a workspace leaderboard, including your best score, ranking, and the score ahead of you. (#953, #976, #981, #991, #1002)

Performance and reliability fixes

Completed work survives task limits

When a task reaches its model or tool-call limit, Hyper now keeps the work already completed in the conversation.

Tools that never started are marked as not run, and Hyper explains that you can continue with another message.

The limits themselves have not changed, and this does not recover work from older failed runs. (#901)

Stopping is faster and preserves your input

The app now shows Stopping immediately, and the server notifies active execution when a stop is committed rather than waiting only for a periodic check.

Messages that do not remain in the transcript are returned to you instead of disappearing. Stopping also clears pending work rather than unexpectedly starting another queued message.

An action already sent to an external service may still complete; stopping Hyper does not undo that action. (#883, #1042, #1047)

More reliable iPhone sending

We fixed cases where sent text reappeared in the composer even though the message remained in the conversation.

You can send another instruction while a reply is running. The send control no longer waits behind the whole task, and pressing Stop does not accidentally release a waiting message as a new turn. (#924)

Webflow connections work again

Webflow sign-in no longer requests an unsupported workspace-level scope that caused the provider to reject consent.

Provider errors are also no longer mislabeled as “Access was declined in the browser.” Hyper distinguishes an actual declined consent from a failed connection attempt. (#1020)

Vapi API keys and synchronization

Vapi now validates private API keys through a supported route and detects whether the key belongs to the US or EU region.

We also fixed chat pagination so a full synchronization does not fail when Vapi omits an optional pagination field.

API-key setup pages now show numbered instructions instead of one dense paragraph. (#1021)

Searchable Notion database properties

Notion database rows often store their useful information in properties rather than in the page body.

Hyper now indexes supported properties as readable text, including statuses, dates, numbers, people, URLs, and supported formula and rollup values.

This makes property-only rows searchable. Existing pages gain the updated content when they next synchronize. (#984)

Cleaner Gmail indexing

HTML-only emails are now indexed as readable text rather than raw HTML.

Gmail background indexing also excludes the Promotions, Social, Updates, and Forums categories. Primary, Sent, and archived messages outside those categories remain eligible.

Existing indexed messages are updated or removed through the normal full-sync cycle, so the change is not immediate for every mailbox. (#970, #971)

Better model and connector error handling

Claude Opus 5.5: Uses standard processing instead of requesting unavailable fast-mode capacity.
OpenAI responses: Preserves the order of interleaved reasoning, text, and tool calls when reconstructing model input.
Connector failures: Retains more useful provider error detail and response headers.
Provider responses: Distinguishes a provider’s refusal from a failure in Hyper’s own execution.
Tool results: Handles null characters that could otherwise prevent results from being saved.

(#880, #889, #900, #973, #975, #977)

Several fixes in this section already shipped in 1.8.1–1.8.4 and are included for anyone upgrading from 1.8.0.

Other additions and improvements

Reliable file drops: Fixed a macOS issue where dragging an image inside a conversation prevented later file drops from Finder. Unsupported dropped HTML is also kept out of the composer. (#1077)
Predictable zoom: App zoom now changes through keyboard shortcuts rather than modified mouse-wheel gestures. Use Command/Ctrl + + or -, and Command/Ctrl + 0 to return to 100%. (#1028)
Planner operation limits: Corrected the exposed Planner interface to exclude edits and deletions that require unsupported version headers while retaining supported creation operations. (#1011, #1015)
More polished branding: Updated the home wordmark, connector artwork, sidebar accents, and light- and dark-mode presentation. (#939, #947, #961)
Smaller desktop downloads: Improved desktop build optimization, removed a development-only tool from the app bundle, and improved Mac disk-image compression. (#897, #911, #913, #929)

Engineering and release improvements

Less unnecessary synchronization work

Unchanged files no longer require the same repeated database rewrites.

We also removed unused indexed-file metadata and stopped retaining successful synchronization jobs indefinitely. Notion properties were moved into searchable content before the unused metadata was removed.

These changes reduce database work and storage while supporting the new search index. (#949, #974, #979, #984, #985, #987)

Better visibility into failures

We added connected tracing across server requests, agent runs, tool calls, and connector requests.

Monitoring now covers failed tasks, stuck jobs, connector health, sign-in and token-refresh failures, and webhook renewal. Shared telemetry code gives these operations consistent outcomes and timings.

Restricted monitoring access and allowlisted trace attributes keep diagnostic access scoped to the information each monitor needs. (#993, #994, #995, #996, #998, #1041, #1043, #1049, #1057, #1073)

Stronger builds and diagnostics

Desktop releases retain matching debug symbols separately from the smaller shipped app. Local app logs also record crashes more usefully.

Build checks now reject additional frontend and Rust compatibility warnings. Release tooling improves Intel Mac builds, change detection, documentation builds, and deployment notifications. (#905, #913, #915, #934, #962, #969, #1014)

Foundations for delegated work

We added durable storage, execution, and controls for child agents, including conversation-owned jobs that use the existing queue and recovery system.

This is groundwork for delegated work. The chat integration is not included as an available feature in this release. (#960, #1025, #1026, #855, #1052)

Coming next

More Hypercards for connected tools: Views such as calendar events, Linear issues, PostHog charts, and boards. The connector-view framework is in place; these example views are the next step, not features included in this release. (#1088, #1092, Hypercards plan)
Browser use: Let Hyper open pages, read, click, type, and work across tabs on your connected desktop. The browser tool is available in nightly builds, with conversation-level approval, but is not enabled in the stable release. (#963, #964, #965, #966, #1033)

Have a feature your team needs? Tell us at founders@heyhyper.ai.