Changelog

Build 90

BoltAI now feels more at home on your Mac. The rebuilt chat window brings a native toolbar and sidebar, scrolling that blends into the window, and faster chat switching. This update covers 2.17.0 through 2.17.2, including a fix for a launch crash on macOS 13 and 14.

On macOS 13 or 14 and BoltAI quits right after opening? Download 2.17.2 and install it manually. Here's how.

New

• A more native chat window: the toolbar, sidebar search, and window layout fit macOS more closely. On macOS 26, chat content scrolls beneath the toolbar and the inspector blends with its header.
• Customize the toolbar with the actions you use most, including Share and one-click PDF export.
• Experimental chat tabs: press Command-T for a new tab, or right-click a chat and choose Open in New Tab. Each tab keeps its own place, and you can reopen closed tabs. To try them, go to Settings > Developer > Feature Flags, turn on "Enable native chat tabs (experimental)", then restart BoltAI.
• Review and remove duplicate starter prompts, agents, and workflows from your library. BoltAI shows what will be removed before you confirm.

Improved

• The reasoning menu next to the model picker now offers Extra High and Max for supported Claude models.
• Kimi K3 can now use Low and High reasoning, not only Max, to keep costs down.
• Supported Claude models have a larger default response allowance, giving them more room to think and finish an answer. Your custom limits still apply.
• Long sessions use less memory, especially with document attachments.
• AI Workflows explain capture and permission problems more clearly, help you turn on Accessibility, and let you review an incomplete result.
• Unavailable MCP tools stay visible in chat, agent, and project settings, so you can remove them. Deleting an MCP server also removes its tool selections, including on your other devices.
• Model pickers keep cached models visible while an updated list loads.
• BoltAI no longer offers to install starter prompts, agents, or workflows you already have.

Fixed

• Fixed BoltAI crashing at launch on macOS 13 and 14.
• Fixed the About window opening by itself at launch on macOS 13 and 14.
• Fixed "Hide Dock icon when last window closes" not working reliably on macOS 13 and 14.
• Project folders in the sidebar show their orange color again on macOS 14 and 15.
• Fixed a crash when stopping a response while it streams, a rare crash in background sync, and a crash when stopping speech playback.
• Fixed AI Workflows that insert or copy their result directly failing with "The model returned an empty response" when using Codex CLI. [ref]
• Fixed chats failing to send when an MCP server could not load. AI Workflows also continue when a selected MCP tool is unavailable.
• Fixed results or errors from an earlier AI Workflow appearing in a later run, and No Preview workflows not completing when run back to back.
• Fixed certain AI Workflow requests failing with custom API base URLs.
• Saved custom OpenRouter endpoints are now used when loading models and sending requests.
• Custom agents keep their selected plugins, and interrupted settings saves no longer show a cancellation error. [ref]
• Claude now reports when its response allowance runs out before it answers. [ref]
• Answers that fail to save stay visible when you switch chats, and regenerating one no longer brings back later messages.
• Documents keep their extracted text if the original file is missing, and chats continue when an attached document can't be read.
• Dragged images use the correct format, and attachment-loading failures show an error.
• Prompt editor headings and buttons stay visible in their sheets.

Note: AI Workflows need Accessibility permission. If your workflows don't run, turn on BoltAI in System Settings > Privacy & Security > Accessibility, then try again.

Build 85

This release brings more capable projects, faster AI workflows, and smoother everyday performance, along with a wide range of bug fixes.

New

• Projects now support linked files and folders on your Mac. Your AI can search, read, and edit them as you work. This feature is in beta. [ref] [ref]
• Choose a default agent for each project, so new chats start with that agent's instructions, model, and plugins. [ref]
• BoltAI can automatically read AGENTS.md, AGENT.md, and CLAUDE.md from your linked project folders, saving you from copying instructions into each project.
• "Reveal in Sidebar" takes you from a search result to the chat's location, so you can continue browsing your chats without the search filter. [ref]
• You can now connect your OpenRouter account via OAuth instead of manually entering an API key.
• GPT-6 Astra is now available through Codex CLI. Grok gains Extra High reasoning, and supported Claude models have more reasoning levels to choose from. [ref]
• Amazon Bedrock now caches prompts automatically to reduce the cost of reusing the same instructions and context, with a choice of cache duration.
• Approve edits to linked files directly in chat, either once or for the project. If you're in another app, BoltAI can notify you when it needs your approval.

Improved

• Redesigned project settings to make them easier to navigate. You can now drag files and folders into the new-project window to link them.
• Improved startup speed, chat switching, and the display of code in responses.
• You can now disable providers you don't use, including Setapp, to keep the model picker easier to browse. [ref]
• Improved AI workflow performance.

Fixed

• Fixed crashes related to clipboard handling in AI workflows.
• Fixed slow sidebar searches in large chat libraries.
• Fixed code highlighting losing its colors while a response was being generated.
• Claude now shows an error when it returns an empty response, rather than leaving you without a reply or explanation. [ref]
• Interrupted Claude responses are now marked as incomplete instead of looking like a finished answer. [ref]
• PDF attachments that can't be read now show an error instead of being silently left out of your request.
• Fixed errors when continuing a Claude chat after it read a web page. [ref]
• Fixed plugin-related errors when your chat settings limit how much conversation history is included. [ref]
• Switching providers no longer sends plugins that the new provider doesn't support.
• Fixed Mistral responses failing as they finished generating.
• Fixed duplicate responses and empty response controls caused by overlapping requests. [ref]
• Fixed follow-up messages being lost when sent just after pressing Stop. [ref]
• BoltAI now explains when an agent stops because it has reached its step limit, including after a plugin fails.
• Fixed the main window staying hidden when you clicked BoltAI in the Dock or switched to it with Command-Tab. [ref]
• Opening Instant Chat or Workflow Palette no longer unexpectedly brings up the main window. [ref]
• Fixed HEIC photos dragged from Apple Photos being rejected by AI providers.
• Refreshing the model list no longer keeps showing an outdated locally cached list.
• Failed image generation now shows an error instead of leaving an empty response or a stuck progress indicator.
• Fixed requests timing out after 60 seconds despite a longer timeout in Settings.
• Fixed chats remaining selected in Trash after they were deleted or restored. [ref]

Build 78

This release focuses on protecting your chats and provider credentials, improving Cloud Sync and backup recovery, and fixing reliability issues across Instant Chat, project defaults, streaming responses, provider models, and macOS conversation controls.

New

• Added Sync Now to the File menu and sidebar settings menu, using the same safety checks as Cloud Sync settings.

• Added a More menu to Instant Chat for quick access to conversation actions.

Improved

• Improved sidebar search responsiveness and ensured hidden or deleted chats are excluded before results are ranked.

• Improved provider credential syncing so related credential data is transferred together and updated credentials take effect without stale cached values.

• Improved Backup & Restore warnings to distinguish unprotected credentials in the active account from those stored in inactive backup accounts.

• Improved MiniMax model availability with current model discovery, reliable fallback models, and the correct MiniMax M3 identifier.

Fixed

• Fixed API keys incorrectly prompting for a Master Password or appearing encrypted after relaunch, while keeping credential access scoped to the correct BoltAI account.

• Fixed legacy folder deletions that could leave deleted chats searchable or allow chats and attachments to return after Cloud Sync.

• Fixed deleted chats briefly flashing in the sidebar after using Delete Local Data and relaunching BoltAI.

• Fixed Control-click not opening native context menus for chats and projects in the sidebar.

• Fixed Save Temporary Chat and other Conversation commands being disabled or targeting the wrong chat in Instant Chat and companion windows.

• Fixed Instant Chat prompts occasionally being sent more than once.

• Fixed OpenRouter conversations losing reasoning details when continuing an existing chat.

• Fixed Project Defaults appearing to save a provider or model while project cards and newly created chats continued using the previous selection.

• Fixed Return while renaming a project also creating a new chat, and improved the alignment of Project Default controls.

• Fixed incompatible Google Deep Research agents appearing as regular chat models and causing unsupported requests.

• Fixed Last Sync appearing stale during normal use or reverting to Never synced after relaunch.

• Fixed restored backups leaving Cloud Sync unable to reconnect. BoltAI now safely downloads current cloud data before offering the normal Local Account merge flow, while preserving restored local data if recovery cannot continue.

• Fixed partial AI responses disappearing when generation reached an output limit, was stopped, disconnected, or ended with a provider error. Partial text and reasoning are preserved, terminal errors remain visible, and Continue is available when the provider reports an output limit.

• Fixed Retry and Regenerate occasionally failing when an interrupted response had not finished saving.

• Fixed empty completed responses from some providers surfacing an unrelated metadata error instead of the actual result.