maestrly.
EN ⌄Download free

Release notes

What's new.

Every Maestrly release, newest first. The latest is v0.15.0, released October 8, 2026.

Download the latestChangelog on GitHub

v0.15.0

LatestRelease on GitHub

Added

  • Create or clone a project directly from a chat and start development in it. Choose a Projects folder in Settings → Execution; GitHub search and repository creation use your signed-in GitHub CLI. New GitHub repositories are private by default, and public publication requires confirmation. See creating or cloning a project.
  • Search a bot conversation with Cmd/Ctrl+F, step through matches with Enter or Shift+Enter, and close search with Esc. Search includes older transcript pages and opens folded activity containing a match.
  • Let a bot of your bot server start and follow development conversations in the projects of your computers. Each computer gives the bot access on its own, in the bot's Settings → This computer, choosing the projects, models, actions and approval ceiling, and lists the other computers with access so you can remove them. Calls reach a computer only through the event stream it keeps open to the server, and fail at once when it is offline. Update the desktop app, the bot server and the bot environment to use it. See work in the projects of your computers.

Changed

  • Place the activity line after the agent's text and cards, and keep its latest text visible while it works. Tool screenshots stay in their details unless the agent explicitly shares them; shared screenshots and generated images remain visible in the conversation.
  • Open bot conversations at the end and offer a go-to-bottom button after scrolling up. Following new messages resumes when you return to the end.

Fixed

  • Preserve an unsent bot message when switching conversations, and keep a sent message visible while the bot prepares its turn.
  • Refresh a bot's context window after changing its model, so its context meter and compaction use the selected model's limits.
  • Keep the embedded VS Code memory snapshot files out of project Git status, including main checkouts, and write snapshots only when the app requests one.
  • Update the image library bundled with the local AI runtime to sharp 0.35.5, which fixes a vulnerability in its SVG renderer (GHSA-wq5f-xc86-pv6w), and the MCP SDK copy used by the Claude runtime to 1.31.0 (GHSA-6qxp-vccf-f47h). The app installs the updated local AI runtime from its own package, without network access.
  • Save an environment's settings when Cmd/Ctrl+S is pressed right after an edit. The shortcut could arrive before the settings panel registered the change and was then ignored, leaving the edit unsaved.

Removed

  • Breaking: remove the connection for external agents such as Grok Bot and other MCP clients; use Maestrly bots instead. Desktop no longer offers Settings → Bots or its /mcp/bots endpoint, and the server no longer serves /mcp, connectors, delegated tasks, or their web pages. Conversations such an agent created remain, with their bot badge marked revoked, and become yours to write in. Upgrading revokes the connections the agents had on this computer and deletes the OAuth clients registered for them.

v0.14.1

Release on GitHub

Fixed

  • Wait for active artifact queries before cleaning up storage from earlier versions, and let new queries wait until that cleanup finishes. This prevents the local host from reopening during removal and reporting that it stopped after artifacts were moved to the bot server.

v0.14.0

Release on GitHub

Changed

  • Breaking: artifacts are hosted only on the bot server. Without a paired server, agents ask you to connect one instead of publishing on this computer, and Settings → Artifacts offers an artifact-only server setup. Artifacts that earlier versions hosted on this computer no longer open until you move them to the server, with their IDs, versions, comments, and previews, or delete them; shared ones become private. See artifacts from earlier versions.
  • Give each bot one Linux desktop with its Maestrly browser, terminal, files, and other programs together. A dock opens its browser, running shells, and home folder; browser logins remain shared by the environment. Update the bot environment to use the unified desktop; older images keep separate Browser and Apps screens. See bot screens.
  • Show the bot's conversation beside its computer, and the environment's overview beside its screen, with resizable panels and a narrow-window switch. Bot and environment settings use tabs and keep unsaved changes together in one save bar.
  • Enable Maestrly tools by default, with separate controls for Terminal, Browser, Notes, Memory, Conversation history, Debugger, Artifacts, and Kanban in Settings and each conversation. Existing explicit opt-outs are preserved.
  • Move server, runner, gateway, and bot containers to Node.js 24 and update dependencies. Contributors now need Node.js 24.11 or later within Node.js 24.

Added

  • Send a development plan from a standalone chat to a registered project workspace on an explicit request, choosing a new branch and its base while preserving model, effort, and Fast settings. Retries reuse the same destination conversation. Ask chats keep their read-only permissions.
  • Open standalone chats, workspace agents, and bot conversations in independent windows. Live responses, drafts, attachments, and queued messages stay with the conversation when you move it out or return it to the app. See separate chat windows.
  • Share a personal memory collection across standalone chats and providers, with search, editing, archive, export, and recall controls. Optional automatic saving starts off; project and bot memories remain separate. Memory stays in this computer's application profile. See personal memory.
  • Attach PDFs and UTF-8 text or code files to bot messages, and download files generated by bots to Downloads without overwriting existing files. Downloaded copies remain available across navigation and restarts, with Show in folder.
  • Manage an environment's accounts, models, skills, tools and MCP servers, runtime components, and preferences from the desktop app. The settings are shared by its bots; update the gateway and restart the environment on the new image to use them. See environment settings.
  • The bot server's own port now also serves the artifact viewer, so one route to the server, such as Tailscale Serve, reaches bots and artifacts. See reach the server from another device.
  • Clicking outside the artifact details panel closes it.

Fixed

  • Retry Claude turns that have not yet produced output when another process is refreshing the same account's OAuth token, instead of signing out the account and interrupting its other conversations. Model and usage discovery also retry this temporary contention.

v0.13.0

Release on GitHub

Added

  • Connect Google AI Pro or Ultra accounts through Google's official Antigravity ACP server and use its Gemini models in chat, Maestro, subagents, and bot environments. The runtime downloads on first sign-in; accounts and bot sign-ins remain separate. See Google AI.
  • Check, update, or roll back the Antigravity runtime in Components. Bots update installed runtimes automatically and switch existing processes when the environment is idle. Failed updates keep the current version and account data.
  • Host artifacts on a paired bot server so shared pages stay available while your computer sleeps. Choose where desktop conversations publish, control publication per bot, and browse both hosts in the Artifacts center. Existing servers and migrated bots start with artifact hosting/publication off.
  • Install an artifact-only server with just the gateway image; creating the first bot downloads its runtime later. Installer updates add the artifact viewer port, with a separate SSH tunnel for VPS setups. Public addresses and Tailscale access remain manual. See server artifacts.

v0.12.0

Release on GitHub

Added

  • Dictate messages locally with whisper.cpp in chat and bot composers. Send transcriptions automatically or review them first, choose the language, and hold the microphone button to record. Recordings without speech are discarded.
  • Experimental: cap a bot's context window in its compaction settings, including the shared environment default, to bound context use and cost.
  • Agents can publish web pages as artifacts, such as prototypes, reports, and dashboards, with Maestrly tools on. Each change becomes a new version. Artifacts, in the sidebar, shows every artifact with a preview of its page, its versions, and the conversation it came from; pages open in an isolated frame. See Artifacts.
  • Share an artifact with other people: personal links that show every device that joins, access requests you approve, or a link for anyone with an optional access code and expiry. You can revoke a device, a person, or every session, and the sidebar counts new activity. Sharing needs a public address that you provide, such as Tailscale Serve. See Sharing.
  • People who can open a shared artifact can comment on a passage or a spot of the page. Comments appear as pins on the page, with previews, conversations beside them, and a list; the viewer's top bar also gains version steps, page widths, and full screen. You read, answer, and resolve comments in Artifacts, agents can read and answer them with Maestrly tools, and Send to conversation puts the open comments in the conversation's message box. See Comments.
  • Experimental: see a bot's reasoning in its conversation, like in chats, once its bot server and environment are updated. Older bot servers and computers keep working without it.
  • Choose how conversations show agent activity in Settings → Appearance & sound → Agent activity: Compact (the default) or Expanded, which shows every reasoning block and tool card as before.
  • Experimental: see when bots can be updated and update them in one click with Update bots. Maestrly updates the bot server it installed, then each environment restarts on the new image once none of its bots is working, waiting for you, or under your control, even while your computer is off. An environment's view shows what its update waits for and offers Update now and Cancel update.
  • Experimental: bots update Claude Code and Codex on their own, without waiting for a Maestrly release. Each environment checks npm every six hours, tests a new release before using it, never interrupts a turn in progress, and never goes below the version its image ships. A bot's Settings → Components shows the versions and offers manual checks; servers can turn automatic checks off with MAESTRLY_GATEWAY_BOT_RUNTIME_UPDATES=off.
  • Experimental: an environment's view shows the Claude Code and Codex versions its bots run, where each comes from, an installed version still waiting for work in progress to end, and Check for updates.

Changed

  • Require macOS 13 or later. Apps on macOS 12 are not offered this update.
  • Fold the reasoning, commands, searches, and other tool steps behind each answer into one activity line in chats, workspaces, subagent transcripts, and bot conversations. While the agent works, the line shows its current step; afterwards it summarizes what it did, including steps that failed or that you denied. Open it to see every step and its details. The answer, questions, plans, subagents, published artifacts, generated images, and tool screenshots stay in view.

Fixed

  • Share plain-text clipboard contents while controlling a bot or environment screen, preserve Caps Lock, and show the bot's pointer on its screen.
  • Keep subagent cards visible outside the compact activity line, and show bots' to-do lists without offering desktop-only tools inside bot environments.
  • Let bots use new Claude models such as Opus 5.5: the bot image now ships Claude Code 2.1.285, so bots see these models as the desktop app does, including the extra high effort, instead of a reduced context window and a request to update Claude Code on every turn.
  • Close only the open list, not the whole dialog, when pressing Escape in a searchable dropdown such as the environment choice when creating a bot.
  • Never move a bot server back to an older version when another computer already updated it past this app.

v0.11.0

Release on GitHub

Added

  • Experimental: run remote bots as separate Maestrly desktops in Docker, on this computer or on your own Linux server, with their own screens, browser, model accounts, and files. Bots on a server keep working while your computer is off.
  • Set up the bot server from Settings → Bot server: Maestrly installs it on this computer's Docker or on an Ubuntu or Debian VPS over SSH, pairs this computer, and updates the server to the app's version. Bots reach only the public internet unless you let them reach private networks.
  • Group up to eight bots in an environment that shares one desktop, home folder, and set of model accounts, skills, MCP servers, and site logins.
  • Pair your computer with a bot gateway you manage, manage bots from the Bots sidebar, watch or take control of a bot's screen, and answer its questions and approval requests from Awaiting you.
  • Add model accounts to an environment or bring them from your computer, schedule routines on the server, allow selected bots to exchange bounded messages, and review their activity.
  • See the screenshots and images a bot's tools produce in its conversation, and send it images from your computer by attaching, pasting, or dropping them.
  • Choose a bot's reasoning effort, fast mode, and access from its composer, dictate messages with the microphone, and follow its context use and estimated cost; its model list follows the models you enabled for its account.
  • Hear an alert when a bot finishes or fails what you asked, or needs you.
  • Split the sidebar into Chats, Workspaces, and Bots tabs.
  • Recall relevant memories for each message in project conversations with memory; a chip under the message opens the recalled memories. Turn it off in Settings → Chat → Memory.
  • Save memories from conversations in the background with a memory model you choose, and consolidate overlapping automatic memories. Saving starts off.
  • Let agents search and read their own conversation history.

Changed

  • Encrypt MCP connection details (URLs, headers, commands, arguments, and environment values) at rest when secure storage is available. Earlier versions cannot read them: after a downgrade, set up those MCP servers again.
  • Open and continue long conversations faster.

Fixed

  • Deliver browser and desktop screenshots to GPT-6 models on ChatGPT subscriptions when they call tools from code.
  • Remove the temporary directory of a Codex runtime check on Windows instead of leaving it behind.

v0.10.0

Release on GitHub

Added

  • Attach PDF files to chat messages. Models that read PDFs receive the document; other runtimes receive text extracted locally.
  • Click a PDF in a sent message to open it in the system's PDF viewer.
  • Drag files onto the chat composer to attach them.

Fixed

  • Delete a conversation's attached images and PDFs when the conversation is deleted, instead of leaving them in the application profile.

v0.9.3

Release on GitHub

Added

  • Send requested tasks to new conversations in isolated worktrees, and hand approved plans to new Standard conversations.

Fixed

  • Run independent subagent tasks in parallel for Claude and Codex conversations.

v0.9.2

Release on GitHub

Fixed

  • Ignore malformed or incomplete task-list tool input in task cards instead of interrupting the chat interface.

v0.9.1

Release on GitHub

Added

  • Update the Codex runtime without waiting for a Maestrly release. Settings › Maestrly Chat › Components checks the latest official stable release, tests its compatibility before switching, keeps the previous version for rollback, and can install updates automatically when enabled. Open conversations keep their current version until Maestrly restarts.

Fixed

  • Update the integrated Codex runtime to 0.155.1 so connected accounts can discover GPT-6 Sol and GPT-6 Luna from the provider catalog.
  • Wait for the native macOS updater to finish preparing a downloaded update before offering the restart, and keep Maestrly running with a retryable error if the installer does not take over shutdown.

v0.9.0

Release on GitHub

Added

  • Connect a personal bot, such as a Grok bot, to the native conversations on your own desktop through an MCP endpoint the application serves itself, with no server, database or relay, and no organization, project, board, card or runner involved.
  • Turn that endpoint on in Settings, choosing the local address it binds to and the public HTTPS address the bot dials; it is off until you enable it, and it answers only requests that arrive under that address.
  • Approve or deny every bot authorization on the desktop itself, grant access per bot and workspace, and revoke it at any time; a bot only ever sees the conversations it created itself.
  • Give every bot conversation a fresh, exclusive worktree, and resume that same worktree when the bot continues the conversation.
  • Show bot conversations in the sidebar with a badge naming the bot, and let the person pause one so every bot instruction is refused until they resume it.
  • Choose how far each bot goes before it asks you — request approval, approve for me, or full access — when you approve its request and later on its card; the bot is offered only the modes at or under that ceiling, and plan approvals always stay with you.
  • Release a bot conversation for your own messages, after a confirmation that says what it means: the bot keeps the conversation and is not told what you write, so ask it to read the conversation again when that matters. Only one turn runs at a time, a bot instruction waits for the turn you started, and a bot can neither cancel nor steer it. Blocking your messages again closes the composer without interrupting anything, and pausing the bot still hands the whole conversation back to you.
  • Let a bot read a conversation of its own back page by page, so it can pick up what you wrote in it; it reads public message text alone, never reasoning, tool work, attachments or any other conversation.

Fixed

  • Keep the composer showing the account, model, reasoning effort, fast mode, permission mode and behavior mode a conversation will actually run with when something other than you changes them — a bot configuring its conversation, a delegated stage, or a provider failover — instead of the previous values until the conversation is reopened.

v0.8.0

Release on GitHub

Added

  • Add standalone chats with private persistent files, isolated permissions, provider support, and sidebar lifecycle controls outside project workspaces.
  • Add native Cursor subscription support across chat, Maestro, subagents, image interpretation, desktop execution, and account-isolated sessions.
  • Add optional incremental background compaction with persistent checkpoints, resume, cancellation, retry controls, and provider-specific configuration.
  • Add in-app update discovery and installation through GitHub Releases on macOS, Windows, and AppImage, with release notifications for Debian packages.
  • Delegate development work to an external agent (for example a Grok Bot routine) through an authenticated MCP endpoint, with per-project and per-action grants the owner controls and can revoke at any time.
  • Choose the account, model, reasoning effort, fast mode and execution mode for each stage of a delegated task, and change them while it runs; an effort is never translated between providers.
  • Follow a delegated task to delivery: independent review bound to the exact code revision, required checks, artifacts, commit, push, pull request and merge, plus follow-up rounds when checks fail or changes are requested.
  • Notify an external routine with a signed, at-least-once callback instead of polling, and manage connectors, delegations and OAuth consent in the web interface.

Changed

  • Redesign Kanban card and column-agent dialogs with responsive layouts, clearer configuration, contextual prompts, and concurrent agent execution.
  • Replace the fixed drawer tool strip with searchable, reorderable, persistent tabs stored per conversation.
  • Update compatible minor and patch dependencies across all workspaces.

Fixed

  • Keep messages visible and queued correctly during compaction, align portable Codex recovery with resolved context settings, and recover short turns whose terminal events are lost.
  • Repair legacy databases during startup so workspace removal succeeds, while preventing unrelated UI updates from interrupting live chat subscriptions.

v0.7.0

Release on GitHub

Added

  • Preserve the complete available conversation when switching models or providers, compacting before dispatch when transport limits require it.
  • Block model transfers that remain oversized or whose required compaction fails, avoiding silent context truncation.

Fixed

  • Restore Kanban project chat tool execution for Claude and avoid repeating project context after the first message.

v0.6.1

Release on GitHub

Fixed

  • Stream live context-window usage for Codex/Astra and Claude instead of showing stale measurements.
  • Show compaction progress and failures, with staged retries that reuse successfully completed work.

v0.6.0

Release on GitHub

Added

  • Give project chat scoped control over boards, columns, cards, comments, and column automations, with versioned and idempotent mutations.
  • Refresh board and open-card details in web clients through live server events without overwriting in-progress drafts.

Fixed

  • Use a properly sized macOS template icon for the menu bar, removing the white square while preserving platform-specific colored icons elsewhere.

v0.5.0

Release on GitHub

Added

  • Link desktop workspaces and conversations to Kanban projects and boards, with scoped read and write tools for local agents, Maestro, and GPT Web.
  • Enforce project permissions, resource versions, idempotency, and audit attribution for agent-driven board mutations.

Changed

  • Replace model-specific harness branching with versioned declarative profiles shared across providers, prompts, sessions, subagents, and live controls.

Fixed

  • Hide the paired review loop badge after the loop reaches a terminal state.

v0.4.3

Release on GitHub

Fixed

  • Compile internal workspace packages before creating desktop installers, fixing the missing module error on startup in v0.4.2.
  • Reject desktop packages that omit compiled workspace entry points.

v0.4.2

Release on GitHub

Fixed

  • Use a Debian-safe package name and the desktop workspace artifact paths in native release builds.
  • Validate native packages before tagging without launching packaged GUI apps in hosted runner sessions.

v0.4.1

Release on GitHub

Fixed

  • Restore native packaging and packaged smoke commands used by the release workflow after the monorepo workspace migration.

v0.4.0

Release on GitHub

Added

  • Multi-tenant Kanban platform with audited execution, project chat, scoped permissions, OAuth Device Flow, and self-hosting support.
  • Opus 5 behavior profile across Claude, Copilot, subagents, resumes, and review executions.

Fixed

  • Shut down delayed Codex processes reliably on Windows.
  • Accept ephemeral Claude sessions in paired review loops and release the split view when a loop reaches a terminal state.

v0.3.0

Release on GitHub

Added

  • Ordered Claude subscription account rotation for chat, subagents, summaries, and image interpretation, with configurable fallback accounts.

Fixed

  • Preserve conversation context and completed tool results when switching subscription accounts after usage limits are reached.
  • Make GPT account failover transparent and preserve physical account ownership.
  • Harden archive extraction and web text parsing.

Changed

  • Update compatible dependencies and development tools.

v0.2.0

Release on GitHub

Added

  • GPT-6 Astra support for OpenAI API and ChatGPT subscription sessions, with conversation continuity and context compaction.
  • Mid-turn text steering and live reasoning updates for subscription runtimes that support them.