Goal for this release
The last release introduced Hypercards, new models, a redesigned desktop interface, and more ways to work across connected tools.
This release helps Hyper take on longer tasks and do more of the work in parallel. It brings subagents, background computer commands, more Hypercards, six new connectors, and improvements to navigation, search, and reliability.
Our goals were to:
- Let Hyper divide larger tasks into independent pieces of work.
- Keep longer tasks going without requiring you to start over.
- Let desktop commands run while Hyper continues with other work.
- Bring calendar events, tasks, messages, charts, and graphs into the conversation.
- Make drafts and proposed changes easier to review before anything happens.
- Expand the tools Hyper can work with.
- Make everyday navigation, editing, and connector use more reliable.
Subagents and longer-running tasks
Hyper can now delegate parts of a task to subagents that work in parallel.
For example, it can research several sources at once, investigate separate parts of a codebase, or work on independent pieces of a larger analysis.
Each subagent receives a specific task and the information it needs. Hyper coordinates the work and brings the results back into the conversation. (#1054)
Independent work runs in parallel
Hyper no longer needs to finish every part of a larger task one after another.
It can start a subagent, continue its own work, and collect the result when it needs it. Subagents can also continue after the initial reply ends.
For example:
Review the changes in this release, check the documentation, and identify anything we should flag before shipping.
Research these companies in parallel and bring the findings together.
Subagents use the same connected-service and computer approval controls as other work in Hyper. Delegating a task does not grant additional access. (#1054, #1164)
Follow progress in the conversation
Subagents appear as compact activity chips that you can open to see their task and tool activity.
Their progress stays with the conversation when you switch chats or continue with another message. When a subagent finishes while the conversation is idle, Hyper can return with a follow-up.
New messages from you take priority over an automatic follow-up. (#1053, #1183, #1207, #1216)
Keep control of delegated work
You can ask Hyper to redirect or stop delegated work.
Pressing Stop cancels unfinished subagents belonging to the conversation, including ones started during an earlier turn.
Subagents cannot create further subagents, use the browser, or ask you questions directly. If they need information, Hyper handles that in the main conversation. Permission requests still use the normal approval flow.
An action already sent to an external service may still complete; stopping Hyper does not undo it. (#1054, #1164, #1165)
Longer tasks can keep going
Hyper can now compact its context during an active task, rather than waiting until the final reply.
When the working context gets large, Hyper summarizes the progress and continues the same task. Your full conversation remains available; compaction changes the information prepared for the model, not the visible conversation history.
Completed tool results and new instructions are preserved while this happens.
We also increased the default limits for new tasks:
- Model calls: 50 → 100.
- Tool calls: 100 → 200.
Existing tasks retain their saved limits. Task-duration and computer-session limits still apply. (#1123, #1125)
Better recovery when context fills up
Hyper now measures compaction against the model’s input limit, reducing cases where a request is rejected for being too long.
If a provider still refuses a request because its context is full, Hyper can compact and retry once.
This does not make tasks unlimited, but it helps longer work continue without losing the progress already made. (#1152, #1155)
Background computer work
Hyper can now start a long-running desktop command and continue with the rest of your task.
You can ask it to:
- Run a build while reviewing the code.
- Start a test suite while investigating another issue.
- Run a local server while preparing the next step.
- Continue other work while a command processes data.
Background commands run on your connected desktop. They are not available in the temporary cloud sandbox. (#1146)
Follow output without blocking the task
When a background command starts, Hyper receives a log path instead of waiting for the command to finish.
Both normal output and errors go to that log. Hyper can read it later to check progress or inspect the result.
A command that did not start has no log. Hyper records each background start so retrying the same request does not launch the command a second time. (#1146)
Background commands have their own lifetime
A background command can continue after Hyper finishes its reply.
It ends when the command exits or reaches the two-hour limit. Ending the chat task does not automatically stop the command.
The same folder and script-approval rules apply to foreground and background commands. Running a command in the background does not bypass approval.
This release does not yet add a dedicated background-command status row or stop control. Hyper checks the command’s log to follow its progress. (#1146)
Larger files from desktop work
Desktop scripts can now return files up to the conversation’s attachment limit.
This removes a smaller computer-output restriction that could prevent an otherwise supported file from being returned. (#1124)
More Hypercards
Hypercards now cover more of the work you do across connected tools.
You can review calendar events, bookings, issues, tasks, messages, emails, charts, and graphs directly in the conversation. Supported actions can appear as editable drafts or change previews before you run them.
The available card depends on the connected service and the operation. If a card cannot show the complete request faithfully, Hyper does not hide the remaining changes inside it.
Draft and review calendar events
Calendar cards support Google Calendar, Outlook Calendar, and Cal.com.
You can:
- View individual calendar events and bookings.
- Draft Google Calendar and Outlook Calendar events.
- Review event details before creating them.
- Draft a Cal.com booking.
- Review supported changes to Google Calendar and Outlook Calendar events alongside their current values.
For example:
Draft a meeting with the team for next Thursday.
Show me what would change before moving this event.
Event and booking cards keep the action in the conversation while you review it. Creating an event, saving a change, or booking a meeting remains a separate step. (#1140, #1143, #1177, #1182, #1225, #1231, #1253)
Some recurring events, rich descriptions, and provider-specific options do not have an editable card form. Cal.com booking cards do not themselves check availability; Hyper must obtain the relevant availability separately.
View and create issues and tasks
Issue and task cards support Linear, Jira, GitHub, GitLab, Asana, Trello, Todoist, and Microsoft Planner and To Do.
Cards show the information available for the item, such as its title, description, status, priority, owner, project, labels, and due date.
You can also draft supported new issues and tasks, edit the available fields, and create the item when you are ready.
For example:
Draft an issue for this bug, with the reproduction steps and expected behavior.
Turn these follow-ups into tasks for me to review.
The fields you can edit depend on the tracker. A GitHub issue card is not a pull-request card, and support for creating a task does not imply that every task field can be changed through a card. (#1136, #1142, #1184, #1218, #1220, #1255)
Review changes before saving
Change cards show supported edits alongside the item’s current values.
This release includes change cards for:
- Google Calendar and Outlook Calendar: Supported event changes.
- Linear: Issue title, description, priority, and due date.
- GitHub: Issue title and body.
- GitLab: Issue title, description, and due date.
- Asana: Task name, notes, and due date.
Choose Save changes when you are ready to apply the edits.
A change card only supports requests whose effects it can show. Changes to other fields, such as an issue’s assignee or state, may need a different approval flow.
Trello, Todoist, Microsoft Planner, Microsoft To Do, and Jira have read and create cards in this release, but not these editable change cards. (#1253, #1258, #1262, #1266)
Draft and read messages in chat
Message cards support Slack, Discord, Microsoft Teams, Telegram, WhatsApp, and LinkedIn.
You can draft a message, edit it in the conversation, and send it when you are ready.
Supported message reads can also appear as cards with the conversation’s messages and available details, rather than a block of raw connector output. Slack cards include channel history and threads.
Slack and Discord also support editable cards for changes to existing messages.
For example:
Draft a Slack update for the team using the decisions from this meeting.
Show me the conversation, then help me write a reply.
Message cards respect each service’s limits. Some attachments, rich layouts, mentions, and other provider-specific message types do not have an editable card form. (#1219, #1224, #1228, #1229, #1257)
More ways to work with email cards
Gmail and Outlook message reads can now appear as email cards.
Supported Outlook replies, reply-all messages, and forwards can also become editable drafts. Successful email actions can show the resulting email card.
This builds on the email compose cards introduced in the last release; it does not replace their existing editing and sending controls.
A Gmail thread card shows the latest-arriving message, not the entire thread as one email. Some Outlook reply requests rely on recipients chosen by Outlook and cannot be represented as a complete editable draft. (#1242, #1257)
Charts and graphs from connected tools
Reports from PostHog, Google Analytics, Grafana, Datadog, and Google Ads can now appear as charts and graphs.
Supported reports include:
- PostHog trends and saved insights.
- Google Analytics reports and real-time reports.
- Grafana data-source query results.
- Datadog metric results.
- Google Ads reporting queries.
Charts preserve the relevant time information from the source. Where a result is incomplete or cut off, Hyper avoids presenting it as a complete time series.
Chart cards are available on desktop and iPhone. (#1159, #1170, #1190, #1257)
Draft when you want to review
Hyper now distinguishes between asking to work on a draft and asking it to take an action.
For example:
Let’s draft a message to the team.
creates an editable card without sending anything.
Send the message to the team.
uses your existing approval settings.
You can also ask Hyper to run a card after reviewing it. Hyper uses the card as you left it, including your edits.
An automation action set to draft still creates a draft rather than running automatically. (#1242)
Edits and results stay with the card
Incoming updates no longer overwrite changes you are typing into a card.
Hyper also keeps the result of a card action, including why an action was not sent, so both you and Hyper can refer to it later.
Where supported, cards show readable names instead of service-specific IDs. If a name cannot be retrieved, the original ID remains available rather than being replaced with a guess. (#1187, #1194, #1208, #1223, #1230)
Six new connectors
Hyper now connects to Salesforce, Figma, Shopify, Cedar, Pinterest, and Grafana.
Supported operations remain limited by the connected account’s permissions, provider restrictions, and Hyper’s action controls.
Salesforce
Connect Salesforce to search and work with CRM records, reports, and other supported data in your organization.
Hyper can query Salesforce, work with supported standard and custom objects, and use reporting, bulk-data, and other supported APIs.
Background synchronization follows record changes and deletions. Search visibility is derived from the source object’s sharing model.
Your Salesforce administrator must install and allow Hyper’s app for the organization. Available operations depend on your Salesforce permissions, and requests whose effects depend on their arguments require review for each run. (#1138)
Figma
Connect Figma to read designs, inspect file and node details, render supported content, and work with comments and other supported resources.
Hyper reads Figma on demand rather than copying design files into a background index.
You can provide a direct Figma file link for Hyper to work with. Team-folder and team-file browsing are not included because the required scope is not available to this integration.
Supported actions include comments, reactions, development resources, and webhooks. Access remains subject to your Figma account and the scopes granted to Hyper. (#1135, #1247)
Shopify
Connect Shopify to work with store data through its Admin API.
Hyper can read and perform supported actions across products, orders, customers, and other store resources, subject to the permissions granted by the store.
Synchronization and webhooks keep supported indexed information current. You can begin installation from Shopify or connect through Hyper without manually entering the store address.
Older order history requires the appropriate Shopify access. Customer names, email addresses, addresses, and other protected data remain subject to Shopify’s protected-customer-data requirements. (#1149, #1217)
Cedar
Connect Cedar to work with deals, tasks, documents, meetings, and other supported sales-workspace information.
Hyper indexes organization deals, your own tasks, and supported shared or personal documents. Other Cedar tools, including mail, calendar, Slack, and LinkedIn operations, are read or used on demand rather than copied into Hyper’s index.
A Cedar connection is private by default because its tools can act on personal accounts.
Permissions distinguish individual actions within a tool. Permission to list tasks, for example, does not also grant permission to delete them. (#1166)
Connect Pinterest to work with boards, Pins, analytics, advertising, and other supported account resources.
Hyper can synchronize boards and Pins, read current data, and perform supported actions such as creating Pins with attached media.
Available operations and limits depend on Pinterest’s access level. Trial access has request limits and restrictions on public Pin publishing; broader production use depends on Standard access.
Changes to advertising or other live account resources remain subject to approval and the connected account’s permissions. (#1145)
Grafana
Connect Grafana to read dashboards, query data sources, inspect alerts, and perform supported dashboard and alert-management actions.
Grafana uses a site address and service-account token rather than an OAuth sign-in.
Hyper reads Grafana on demand. The service account’s role determines which data and actions are available.
The Grafana address must be publicly reachable and cannot include a custom port or sub-path. Server administration, licensing, sign-in configuration, and token-management routes are not exposed.
Disconnecting makes Hyper forget the token. To revoke the token itself, delete it in Grafana. (#1156, #1161)
Existing connectors can do more
Firebase Authentication users
The Firebase connector now includes Firebase Authentication.
Hyper can read registered users, last-sign-in timestamps, and supported tenant accounts, including paginated results.
Account queries are treated as reads. Supported account creation and deletion operations remain subject to write approval.
This uses the existing OAuth scope and preserves the connector’s Firestore support. We also fixed routing for Firebase account creation. (#1119, #1121)
Datadog write actions
Datadog now supports write actions as well as reads.
Supported actions include creating or updating notebooks, monitors, cases, comments, and other supported resources.
Your Datadog administrator must enable MCP write access for the organization. Write operations use Hyper’s approval controls, with per-run review for operations whose effects require it.
We also corrected Datadog’s connection setup so tool calls begin with the required MCP session. (#1197, #1214)
Easier navigation
Preview chats and automations
You can preview chats and automations from the sidebar without opening each one.
Chat previews include a short summary and current activity, making it easier to find the conversation you want and see which work is still running.
Previews update as the conversation changes and remain subject to the same access controls as the conversation itself. (#1205, #1206)
Scroll between earlier replies
Mouse-wheel scrolling can now move between earlier and later turns in a conversation.
You can browse previous replies without relying only on clicking the navigation rail.
Keyboard and tab navigation also include automation runs, while navigation inside an automation remains scoped to that automation. (#1144, #1150)
Clearer tool activity
Tool activity now uses clearer action labels and more useful parameter previews.
You can see the relevant details of a request without the activity row being filled with internal routing fields. Labels remain readable after an action finishes or fails and when you reopen the conversation. (#1186, #1212)
Invite teammates without leaving your work
Workspace administrators can invite members from a popover without leaving Home, a conversation, Search, or Settings.
Settings controls have also been reorganized into more focused categories. (#1134, #1074)
Daylight mode
Choose Daylight in desktop Appearance settings to have Hyper’s colors follow your local time.
Daylight uses the regular Light theme at noon and warmer colors at night, with gradual changes between them. Your selection is saved when you close and reopen the app.
This changes Hyper’s appearance, not your system display settings. On iPhone, Daylight currently uses the light appearance. (#1251)
Connect when you are ready
Hyper can now suggest a useful connector without stopping the conversation.
A suggested connection card stays available after the reply ends, so you can connect the service when you choose. It does not mark the conversation as waiting for you.
If you have already connected the service by the time you return to the card, Hyper uses that state rather than asking you to add a duplicate account.
This is separate from a connection required to continue a particular task, which can still wait for you. (#1263)
Claude Haiku 5.5
We upgraded:
- Claude Haiku 4.5 → Claude Haiku 5.5
The new model replaces the previous Haiku version in the model catalog and relevant smaller-model tasks, including Anthropic subagents.
This update also includes the model’s larger supported context and output limits. (#1191)
A new 3D rocket game
The rocket game now has a 3D chase view and a boost.
You can play while Hyper works, then return to the conversation without stopping the task.
The update also includes the corresponding score-handling changes. (#1189)
Performance and reliability fixes
Give Hyper new instructions while it waits
Messages sent while Hyper is waiting for an answer now reach the waiting task correctly.
You can clarify or redirect the work rather than having the message remain queued or appear to be lost.
Clicking an empty part of the conversation also focuses the composer without taking clicks away from controls or text selection. (#1160, #1153)
More reliable delegated work
We fixed cases where waiting for a subagent or resuming its work could produce duplicate activity rows or leave the composer in a stale state.
Subagent activity stays attached to the correct conversation and task. Internal child transcripts, reasoning, and intermediate tool data remain separate from the progress shown in the main conversation. (#1165, #1188, #1199)
Slack sign-in and reconnects
We removed an unsupported Slack scope that caused Slack to reject the connection request.
When an existing connection lacks a required scope, Hyper can now distinguish operations it can safely skip from cases that need a reconnect.
This avoids treating every missing permission as the same kind of synchronization failure. (#1162, #1176)
Clearer connection status
Connector settings can now show when a connection has stopped synchronizing and when reconnecting is needed.
This helps distinguish a healthy connection from one whose indexed information may no longer be current.
The sync-health changes were briefly backported to 1.9.6 and then removed from that patch. They are included here as part of the new release’s main implementation, not as a feature already available in 1.9.6. (#1171, #1176, #1200)
Better HubSpot, LinkedIn, and Attio synchronization
- HubSpot: Search and batch-read requests are correctly treated as reads, allowing synchronization under read-only grants.
- LinkedIn: A refused post-list request no longer ends the entire synchronization. Other supported data can continue to sync.
- Attio: Recent meetings are read from a checkpoint rather than repeatedly rereading the full workspace.
These changes reduce unnecessary work and prevent one unsupported operation from blocking unrelated information. (#1196, #1227, #1169)
Search stops when its request ends
Database searches now have a time limit aligned with the search tool.
A slow search no longer continues consuming database resources after the request has already timed out. (#1139)
More reliable HTML and card content
We improved how HTML is converted into readable content across connectors.
Malformed markup is less likely to lose text, list and image descriptions are handled more consistently, and deeply nested content can fall back safely.
Email cards also avoid offering an editable form when the original HTML cannot be represented without losing meaningful content. (#1257, #1267)
More reliable shutdown and session recovery
We fixed a crash that could happen when Hyper closed while background work was finishing.
When a saved sign-in session is no longer valid, Hyper now asks you to sign in again rather than leaving the app in a broken session. Temporary rate limits do not trigger the same behavior. (#1126, #1240)
Several improvements in these notes already shipped in 1.9.1–1.9.6, including Firebase Authentication access, Slack sign-in fixes, active-task compaction, the shutdown fix, and Haiku 5.5. They are included for anyone upgrading from 1.9.0.
Other additions and improvements
- Fewer unnecessary files in chat: Files returned by connectors appear in the conversation when Hyper explicitly shows them, rather than every internal file result becoming an attachment. (#1129)
- More reliable Figma consent: Removed an unavailable scope that prevented authorization. The connector no longer offers the folder-browsing routes that require it. (#1247)
- Refreshed welcome screens: Updated the launch animation and return from connector or account authorization. (#1213, #1246)
- Windows installer signing: Official Windows release builds now sign and verify installers before publication, including the publisher identity and timestamp. This does not guarantee that Windows will never show a reputation warning. (#770, #1118)
- Webflow documentation: Added a connector guide explaining setup and supported access. (#1201)
Engineering and release improvements
Less unnecessary synchronization work
Several updates to different items can now be combined into a bounded list of items to refresh, instead of repeatedly triggering a full connector synchronization.
An invalid item can be skipped where appropriate while the remaining items continue.
This reduces repeated work when a connected service sends several notifications close together. (#1148, #1157)
Better visibility into long-running work
We added more detailed monitoring for model calls, computer sessions, and computer actions.
The new records distinguish provider failures, execution failures, timing, and usage. Monitoring dashboards cover model providers, computer use, and database health.
These additions help diagnose slow or failed work without relying only on the visible chat transcript. Monitoring collectors and dashboards require their corresponding deployment updates. (#1180, #1204, #1209, #1210, #1211, #1232)
Better database recovery
We added alerts for slow database queries and stuck connection pools.
The server can also restart when its database pool remains stuck, allowing it to recover rather than continuing to accept work it cannot complete. (#1132, #1133)
More precise connector permissions
MCP tools that support several actions can now classify those actions separately.
For example, reading, creating, updating, and deleting through one provider tool no longer need to share the same permission classification.
Calls with missing or unclassified actions are refused before being sent. Existing connectors retain their established permissions unless explicitly updated. (#1166, #1214)
Stronger builds and release checks
Release tooling now:
- Checks combined merge-queue changes for compilation errors.
- Creates a draft backport PR when a stable cherry-pick has conflicts.
- Builds the docs site without depending on generated app bindings.
- Uses a pinned Supabase CLI for server releases.
- Supports more noninteractive development commands.
- Produces smaller development builds while retaining source-line backtraces.
These are engineering and release changes, not additional settings users need to configure. (#1114, #1192, #1193, #1226, #1244, #1265)
Foundations for more Hypercards
We added shared record-card types for structured fields such as text, numbers, dates, money, choices, and links.
This is groundwork for richer record views and editable forms. Provider-backed CRM record cards are not included as an available feature in this release.
The shared card framework also now handles more read results, write results, and background-task updates consistently across the app. (#1236, #1250, #1252, #1264)
Coming next
This week was a bit slower on new features because we’ve been building the infrastructure to support some massive changes.
- Cloud browser use: Let Hyper open pages, read, click, type, and work across tabs in its own cloud browser without needing your desktop. The cloud implementation is available in nightly builds but is not enabled in the stable release. (#1167, #1172, #1173)
- Google Search Console: Search-performance reports, URL inspection, and supported site and sitemap actions. The connector remains in preview. (#1130)
- A fresh coat of paint: We’re revamping the whole interface, including Hypercards, to make Hyper a phenomenally beautiful experience.
Have a feature your team needs? Tell us at founders@heyhyper.ai.