ChatShell Download

Changelog

What changed in each ChatShell desktop release. The app updates itself, so you get these as they ship.

  1. 0.12.0 Latest

    Added

    • Dark mode. Settings > Appearance offers Light, Dark, and System themes. System is the default and follows the OS appearance as it changes.
    • The assistant can ask you questions mid-reply, with single or multiple choice options, a recommended answer, and a free-text box. Skip a question or choose to chat about it instead. Turn it on or off as a built-in tool.
    • MCP servers can ask for input during a tool call. A form opens next to the approval card, and URL requests open the browser only when you click.
    • MCP prompts appear in the prompt picker. Fill in their arguments and the rendered text is inserted into the message box for review.
    • A new tool lets the assistant read MCP server resources. Large text is trimmed and binary contents are saved with the conversation’s tool output.
    • MCP servers that ask for workspace roots get the conversation’s working directory, or its scratch folder when none is set.
    • MCP tool calls show server progress on the running tool card, and the tool list refreshes when the server reports a change.
    • MCP server settings add a tool-call timeout and a Trust tool annotations switch, off by default. When it is on, tools the server marks as read-only run without asking.
    • An experimental switch per MCP server to use the 2026-07-28 protocol. Servers that do not support it fall back to the earlier version.
    • Parameter presets have an Additional Parameters (JSON) field for provider-specific fields, such as {"reasoning_effort": "low"}.
    • A Review Model picker in Settings > Auxiliary Models chooses the model that reviews actions in Auto mode.

    Changed

    • MCP tools that the server marks as destructive always ask for approval and show a warning on the approval card.
    • Expired MCP sign-ins are refreshed automatically. When a server needs you to sign in again, a banner in the chat offers to re-authorize, and the server shows a Sign-in expired badge.
    • Selected MCP servers connect at the same time instead of one after another.
    • MCP tool results show structured content, and tool calls that run too long end with a timeout error instead of hanging.

    Fixed

    • Conversations with a deleted model keep its avatar and name.
    • Clearing a preset’s additional parameters now removes them instead of keeping the old value.
    • Images returned by tools now reach the model on OpenRouter, Ollama, Azure, and other Chat Completions-style providers.
    • The question and MCP input dialogs keep their title and buttons in view when the form is long.
  2. 0.11.0

    Added

    • Queued messages have a Send now action that stops the current reply and sends that message right away, ahead of the rest of the queue.
    • Contacts open a profile page before starting a chat, with details, badges, and recent conversations. Double-click or press Enter to start a chat right away.
    • Prompts in the Library open a detail page with the full content, the conversations that use it, and copy, star, edit, duplicate, and delete actions. Use in New Chat sets a system prompt or fills the message box with a user prompt.
    • A Presets tab in the Library to create, edit, duplicate, delete, and set the default model parameter preset, with each parameter optionally left to the provider default.
    • Assistants can pick a parameter preset, and the selected preset is now actually sent to the model.
    • Skills move from Settings to Library > Skills, with a detail page showing SKILL.md, files, and status badges. Settings keeps skill sources, rescan, and enable or disable all.
    • ChatShell’s own skills sync across devices together with their files.
    • Enable All and Disable All apply only to the skills shown when a search is active.
    • The sidebar search filters the current list in Library and Contacts.

    Changed

    • The New Conversation button sits at the top of the conversation list and shows its shortcut. An icon button appears in the header while the list is collapsed.
    • The working directory has its own control in the chat toolbar instead of living in the conversation settings menu.
    • Anthropic models are fetched from the Anthropic API, and new providers no longer start with a built-in model list.
    • ChatShell’s own skills are stored in the app data folder, and the skill installer installs into the shared ~/.agents/skills. Skills in ~/.chatshell/skills are moved over automatically.
    • Sent images, files, and links, including those of a resent message, appear immediately instead of after the backend stores or fetches them.
    • Untitled conversations read “New conversation”.

    Fixed

    • Starring an assistant no longer wipes its role, description, user prompt, parameter preset, and image avatar.
    • Removing a model asks for confirmation, and contact menus that did nothing are now wired up or hidden.
    • Skills are no longer deleted for every device when one device lacks a skill’s folder.
    • Enable All and Disable All no longer mark every skill as downloading.
    • Skills marked as hidden from the model are no longer offered to it.
    • Image attachments work with every OpenAI-compatible provider, such as MiniMax.
    • The thinking card collapses once the answer starts streaming, and scrolling up in it no longer jumps back to the end.
    • The @-mention popup is no longer clipped by the message box.
    • Cmd/Ctrl+N keeps working after the first use.
    • Leftover translation conversations no longer show up in the conversation list or sync to other devices.
    • Quitting during sync no longer crashes, and shutdown no longer hangs on blocking work.
  3. 0.10.2

    Changed

    • The sent message appears in the conversation immediately instead of after the send completes.
    • When Auto mode is unavailable, it stays in the permission mode menu as a disabled option with a plain explanation, such as a bad key, a missing model, or a rate limit. The provider’s raw error is shown on hover.

    Fixed

    • Replies to a conversation that runs in the background now finish correctly after switching to another conversation, instead of staying stuck in the streaming state.
    • Sending in one conversation no longer disables sending in every other conversation.
    • Switching back to a conversation that is still replying no longer hides the reply until the next reload.
    • The “Waiting for response” timer keeps counting across conversation switches instead of restarting at 0s.
    • A failed send no longer leaves the conversation stuck waiting for a reply.
    • A queued message that fails to send is kept at the head of the queue with the error and a Retry action, instead of being lost.
    • Translation output is kept intact around think tags, HTML, and model control tokens.
  4. 0.10.1

    Changed

    • Line charts fit their value axis to the data instead of starting at zero, so small movements stay visible. Bar and area charts keep a zero baseline.
    • Chart lines are drawn as straight segments between points, and single-series line charts label their first and last values.
    • Charts can show a unit, such as a currency or a short suffix, on axis ticks, tooltips, and point labels.
    • Line and area charts with dated labels space points by date and show month, year, or day ticks. Tooltips show the full date.

    Fixed

    • Claude models now show their reasoning, and their replies are no longer cut short at 8,192 tokens.
    • Replies cut off at the output limit are saved and marked as cut off instead of ending silently.
    • Auto-review works with Claude models that reject disabling thinking, and retries with a larger budget when its verdict is cut off.
    • A long reason for Auto mode being unavailable no longer stretches the permission mode menu across the window.
  5. 0.10.0

    Added

    • Auto mode. Tool calls run in the OS sandbox (macOS Seatbelt, Linux bubblewrap) and an auto-review model judges each call before it runs, with an option to run a blocked call anyway. Reviews of a turn’s tool calls run together and show as collapsible cards with outcome, risk, and rationale.
    • Ask mode also runs bash in the sandbox where one is available, with a single approval card when a command needs to reach the network or write outside the allowed folders.
    • Remembered approvals. Choose “Remember on this device” on an approval card, and review or clear what you’ve remembered in Settings.
    • A new Permissions section in Settings with the default permission mode, remembered approvals, bash environment variables, and this machine’s sandbox status.
    • Charts and KPI cards in assistant replies: bar, line, area, pie, radar, and radial charts, drawn in a shared color palette with figures and Mermaid diagrams.
    • Each assistant message lists the files it created or modified, and you can undo a turn’s file changes.
    • Type @ in the message box to mention and attach a file from the working directory.
    • Attach videos, CSV and TSV data files, and folders (as a one-level listing).
    • Scanned PDFs can be attached, and the assistant can read their pages as images.
    • Translation history, kept on this device only, with search, restore, and a toggle to stop saving.

    Changed

    • New conversations default to Auto mode on fresh installs (Ask when the model can’t review actions), and switching to Yolo always asks for confirmation.
    • Images are resized to fit provider limits when attached, and oversized files are rejected with a clear reason.
    • Text attachments too long to send in full include their beginning, and the assistant can read the rest on demand.
    • Long command output stays readable in later turns instead of disappearing when the reply ends.
    • The assistant keeps deliverables in your working directory and scratch files in a temporary folder.
    • Attachments of deleted messages are cleaned up on this device and in the cloud.

    Fixed

    • The assistant can no longer read ChatShell’s own data folder, and it asks before reading shell configuration files.
    • Retrying or regenerating a reply sends the original attachments and fetched pages again.
    • Images read by tools reach the model as images rather than text.
    • Dollar amounts in replies no longer render as math.
    • Translation output only follows new text once you scroll to the bottom.
    • Overlapping avatars in the sidebar no longer show a halo around them.
  6. 0.9.2

    Added

    • When an update is available, Settings shows what’s new in it, with a link to the full changelog.
    • Inline SVG figures in assistant replies, with animation controls, zoom, source view, copy, and save.
    • On-device language detection for translation, with Enter choosing the target language based on your input.

    Changed

    • Translation has primary and secondary language buttons, defaults based on your system languages, and shortcuts for translating and copying results.
    • Cross-device sync responds to cloud changes, window focus, and system wake, and sends pending changes when the app loses focus, sleeps, or quits.
    • A new macOS app icon designed for macOS 26, with a fallback for earlier versions.

    Fixed

    • Markdown HTML is sanitized, Mermaid uses strict security, and a content security policy restricts scripts and embedded content.
    • Adding a provider no longer retains the custom provider you were editing.
    • The provider list in Settings scrolls fully, keeping Add Provider accessible.
    • Provider and model avatars keep their original brand colors and fit inside circular frames.
    • The macOS installer window uses the intended icon sizes and layout in release builds.
  7. 0.9.1

    Fixed

    • Links and files open again from inside the app.
    • The macOS build is signed with its iCloud entitlements, so cross-device sync can read the iCloud container.
  8. 0.9.0

    Added

    • In-app updates. ChatShell checks for new versions, verifies every package with minisign, and installs only when you click.
    • Tool approval. Choose between asking before each tool call and letting the assistant run tools without asking.
    • Conversations are now a message tree. Edit or resend a message to branch, then switch between versions of a reply.
    • A welcome screen and plainer first-run copy.
    • Translation moves to a full-page split screen.
    • Failed and waiting turns are shown in the transcript with a retry button.
    • Conversation rows show a generating indicator while a reply streams.
    • Bing as a web search engine, with the default engine picked from your time zone.
    • Refreshed provider catalog and brand icons.

    Changed

    • A new ink-black visual identity across the app.
    • New conversations stay drafts until you send the first message.
    • The sidebar starts collapsed to the icon rail.
    • Each run’s process steps are folded under a single header.
    • More predictable transcript scrolling while replies stream.
    • Switching conversations feels faster.

    Fixed

    • A provider setup that cannot connect can now be repaired instead of blocking the app.
    • Concurrent tool approval requests queue instead of overwriting each other, and grants are cleared when you switch modes.
    • Sync backs off when the iCloud container is congested.
    • Messages can no longer be sent while a reply is still streaming.
    • Office documents preview through text extraction.
    • Web fetch works again with the macOS WebView and with Obscura behind fake-IP proxies.
    • Opening an old conversation no longer moves it to the top of the list.
  9. 0.8.3

    Added

    • Press Up and Down in the message box to browse your earlier prompts.
    • Office documents can be attached and sent to models that read them.
    • Search results are handed to tool-capable models on demand instead of all at once.
    • The read tool can read a page range of a PDF.
    • Web fetch falls back to a better fetcher when a page comes back empty or broken.

    Changed

    • Long conversations scroll and stream much more smoothly.
    • Large attachments are loaded on demand instead of being sent inline every turn.

    Fixed

    • Screenshot export captures the whole conversation.
  10. 0.8.2

    Added

    • File edits made by the assistant are shown as grouped changes.

    Fixed

    • Repeated context compaction in long conversations is more accurate and survives restarts.
    • Clearer error messages when an edit is refused because the file changed.
  11. 0.8.1

    Added

    • Text-only models can take image input through a vision fallback model.
    • Log storage management with size caps, plus an option to have AI diagnose a log.
    • Settings show the version of the model capabilities database.
    • A hint for installing Obscura in the web fetch settings.
    • A redesigned About section.
    • A splash screen while the app starts.

    Fixed

    • Links, file dialogs, and file access work again.
    • Startup no longer stalls while sync keys load.
    • Sync onboarding handles a missing iCloud account and no longer leaves you stuck.
  12. 0.8.0

    Added

    • End-to-end encrypted cross-device sync through iCloud, with the key kept in iCloud Keychain.
    • Sync setup as the last step of onboarding.
    • Sidebar translator, with a button to continue the translation as a conversation.
    • Auxiliary models: route small tasks to a tiny model and image questions to a vision model.
    • Context compaction keeps long conversations within the model’s window.
    • Web fetch uses the system WebView, Obscura, or Chrome, depending on what a page needs.
    • Richer rendering of tool results.

    Changed

    • Updated model capabilities data, refreshed automatically.
    • The sync passphrase flow matches the iOS app.

    Fixed

    • Web search uses the provider you configured.
    • Attachments repaired after a fetch display correctly.
    • Ollama providers stay visible after auto-import during onboarding.
  13. 0.7.1

    Added

    • Export a question and its answer as a screenshot, including while the answer is still streaming.

    Fixed

    • Follow-up turns with OpenAI Responses API models keep their tool call history.
  14. 0.7.0

    Added

    • A security policy that blocks dangerous shell commands and sensitive file paths for the built-in tools.
  15. 0.6.4

    Changed

    • Shell commands wrapped in harmless helpers are now checked by the security policy too.
  16. 0.6.3

    Fixed

    • Resending a message keeps its original attachments.
  17. 0.6.2

    Added

    • Paste a JSON config to import MCP servers.

    Changed

    • Prompt caching is turned on for Anthropic models, which lowers the cost of multi-turn conversations.

    Fixed

    • Environment variables for STDIO MCP servers are handled correctly, with clearer error details.
  18. 0.6.1

    Changed

    • Conversations are sorted by the time of their last message.
    • Reasoning and tool call history is kept when you continue a conversation.

    Fixed

    • Skills with the same name from different sources no longer conflict.
    • The copy button copies messages that contain tool calls.
    • The message box fills the input area after resizing.
    • Several tool calling fixes for streaming reasoning, tool call IDs, and tool outputs.
  19. 0.6.0

    Added

    • Queue messages while a reply is still streaming.
    • Skills from several source directories, with search.
    • A built-in skill installer.
    • Skills are rescanned when you start a conversation or open settings.

    Fixed

    • Stopping a reply also stops its running shell command.
    • Line numbers are no longer selected when you copy from a code block.
    • Long model names no longer break the provider settings layout.
    • Saving a new provider keeps the form and confirms the save.
  20. 0.5.1

    Added

    • Model capability awareness from bundled models.dev data.
    • Image generation with Gemini 3.1 through OpenRouter.
    • The system prompt includes basic environment details such as the OS and date.
    • Linux builds.

    Fixed

    • Images cannot be pasted or dropped for models without vision support.
    • Duplicate messages and images during streaming.
    • Removed skills no longer linger in the list.
  21. 0.5.0

    Added

    • Full-text search across your chat history.
  22. 0.4.12

    Fixed

    • Duplicate messages at the end of a reply.
  23. 0.4.11

    Added

    • Interface translations.

    Fixed

    • Reasoning display for MiniMax and OpenRouter models.
    • OAuth sign-in for MCP servers works again over HTTPS.
    • Tool call previews.
  24. 0.4.9

    Fixed

    • On macOS, tools find commands on your shell PATH when the app is opened from Finder or the Dock.
  25. 0.4.6

    Added

    • A kill_shell tool, persistent shell sessions, and unified tool icons.
    • Edit any model from its actions menu.

    Fixed

    • Stopping a reply keeps the tool calls and reasoning produced so far.
    • Provider API errors show the real details instead of a generic message.
    • Duplicate tool names from lazily loaded MCP servers.
  26. 0.4.5

    Fixed

    • The stop button cancels running tool calls immediately.
    • Only HTTP MCP servers are probed, and STDIO errors show the full cause.
  27. 0.4.3

    Added

    • edit and write tools, and the read tool can read images and PDFs.
    • MCP server connection status and tool lists in settings.
    • A prompt to sign in again when an MCP OAuth token expires.
    • A connectivity check for provider APIs.
    • MCP tool schemas load on demand.

    Changed

    • Updated default models for each provider.
    • MCP sign-in tokens are stored in the encrypted database instead of the OS keychain.

    Fixed

    • Anthropic provider compatibility.
    • The MCP refresh button reconnects the server.
  28. 0.4.1

    Added

    • Fork a conversation from any assistant message.
    • Resend your own messages.
    • OAuth 2.0 and bearer token sign-in for MCP servers.
    • Export messages and conversations as screenshots.
    • Consecutive tool calls collapse into a group, with icons by tool.
    • Reasoning display for models that stream reasoning_content or <think> tags.

    Fixed

    • Inserting a saved prompt places it at the cursor instead of replacing your text.
  29. 0.4.0

    Added

    • Every LLM provider supported by rig, plus MiniMax.
    • Custom OpenAI-compatible and Anthropic-compatible providers.
    • Add models by hand when a provider’s model list cannot be fetched.
    • Turn tools, skills, and MCP servers on or off in bulk.
  30. 0.3.0

    Added

    • Skills: reusable capabilities you can attach to assistants and conversations.
    • Tools, built-in and MCP, can be configured per assistant.
    • A working directory per conversation for the bash tool.

    Changed

    • URLs in pasted text are no longer fetched automatically.
  31. 0.2.0

    Added

    • MCP support over HTTP and STDIO, with tool calls across multiple turns.
    • New conversations inherit your globally enabled MCP servers.
    • web_search and web_fetch as agent tools.
    • Reasoning, tool calls, and text are shown in the order they happened.
  32. 0.1.1

    Added

    • First public release: a desktop AI chat app for your own providers, including OpenAI, Anthropic, Gemini, OpenRouter, and Ollama.
    • Web search and web fetch, with a headless browser fallback.
    • File, image, and web page attachments by paste, drag and drop, or URL.
    • Assistants and a prompt library.
    • Markdown rendering with syntax highlighting and Mermaid diagrams.
    • API keys encrypted at rest.