Your models.
Your projects.
Your space to build.

• CodeSpace Live is now GalaxyDesk. The same Rust-powered workspace, your existing projects and chats, and a new home at galaxydesk.app.
• One floating workspace pane brings Browser, Outputs, Sub-Agents and Inputs together. Switch tabs or dock it on either side, above or below your conversation.
• Shared-browser docking moves more smoothly, keeps the main window in place when you switch chats, and uses a separate browser runtime on macOS.
• Project usage reports show dated model totals, input and output rates, and estimated costs. Export the complete report as CSV or save it as a PDF.
• Projects expand and collapse immediately without rebuilding the whole workspace.

Cost estimates use recorded input and output token counts where available. Codex and Claude Code amounts represent comparable API value, not charges to your subscription; unknown costs remain unpriced.

• Model costs and context sizes refresh at startup and during update checks, with an offline copy when the network is unavailable. Search, sort and sliders help you find the context and price you need.
• A rounded shared browser docks left, right or wide above or below the chat. Opening it widens the app when space allows. Drag the header toward a thin glowing edge, or use the position controls.
• Browser close responds immediately, and steering reaches the active model before the app prepares the larger workspace update.
• Skills and Memory counts sit above their icons as compact badges.
• Private Git history can create its initial checkpoint even when source capture is partial. Subsequent partial captures keep earlier recorded files.

A signed small update. Your model context, saved work and access settings stay under your control.

• Fixes repeated “failed to load workspace requirements” errors with newer Codex versions. The selected project permission profile now remains available when Codex refreshes its session configuration.
• Your access settings continue to apply to every tool call. Existing Codex sign-in, chats and projects are retained.

Available as a signed small update for the 0.8.4 desktop launcher. Fresh installations use the 0.8.4 installer, then receive this update through the same channel.

• Mistakes were made, that's on us. An artificial backup limit could defer updates as saved history grew. This launcher repair removes file-count and total-size quotas from recovery backups.
• Backup inventories now stream to disk with bounded memory. Integrity and path checks remain, and interrupted copies are cleaned up safely on the next launch.
• Recovery now protects private local Git history too, while keeping older backup formats readable.
• A new GalaxyDesk loading screen uses your saved colour scheme before its first frame, with restrained animation and clear startup status. The same palette continues into the app.

This repair changes the desktop launcher and requires the full 0.8.4 installer. Your chats, settings and projects are retained. Future compatible updates continue through the same update channel.

• Model changes carry recent recorded thinking and tool activity alongside the conversation, adapting from up to five handoff sets toward two to fit the selected model. Saved references recover clipped detail.
• Built-in system instructions guide permanent project memory, report creation and lifecycle, verified completion, Git and automatic Multi-Model delegation. Reporting and debugging skills are more focused.
• One Git tool routes to private native source history or the enabled provider linked to the project, with clear limits and existing access controls.
• The desktop conversation and input reserve output-pane space, so opening and closing outputs no longer shifts them.
• Claude Code Full Access automatically approves valid native permission requests, including new tools and plan transitions. Real questions, stopped actions and disabled controls retain their handling.

Handoffs contain reasoning actually returned by the provider, not inaccessible internal thoughts. Project memory remains local and editable; selected context is sent to the model you choose. Compatible desktop installations receive a signed in-app update.

• Local semantic project memory finds earlier decisions and original evidence across project chats and model changes.
• Inspect, edit, disable or remove learned project memories. Global instructions remain yours to manage, and the app shows which project entries were supplied to the current chat.
• Your assistant retains observed executable locations, package information and specific environment failures. Changed environments trigger rechecks; an unchanged known failure can be avoided.
• Skills and Memory share a compact sidebar row with icons and matching count badges. Manage skill descriptions, toggles and editable custom instructions in the Skills window.
• Claude image tool results appear as clickable saved thumbnails alongside other visual evidence.

Inspired by MemPalace’s preservation of original evidence. Storage and search run locally; selected context is sent to the model you choose. Available as a signed in-app update for the current desktop installation.

• Mistakes were made, that's on us. The 0.8.0 update could fail before opening a window when a larger saved history exceeded the launcher's backup-index limit. This release corrects that limit and preserves the working version if a recovery copy cannot be prepared.
• A visible starting window now explains workspace and update preparation. Startup errors stay visible long enough to read.
• Update downloads show their known size, whole-number percentages, and a clear verification stage.
• The Outputs and Subagents panel stays above the message input, with one scroll area instead of clipped sections and competing scrollbars.

Install the full 0.8.1 version to receive the corrected launcher. A small update cannot replace the launcher already installed on your computer. Your chats, settings and projects stay in place.

• Task, Bug and Feature cards bring clear types, icons and filters to each report’s Kanban board. Debugging controls keep reproduction evidence, fixes and verification together.
• Native Rust local Git history captures source checkpoints before and after chat work when no enabled Git provider is linked. Review commits and earlier file contents without installing Git. Existing repositories stay intact.
• Enable Local access to check progress on your phone. Your animated desktop pet and live chat bubbles come along, with projects and chats in a slide-out. The desktop pet setting applies on mobile too.
• Pair a phone or issue a scoped API token for your own tools. Read-only monitoring is the default; optional chat controls retain desktop access and approval rules.
• Models can close, reopen and delete reports and delete individual cards through the project tools.

Local access is opt-in and requires the desktop app to remain open. LAN access uses HTTP on your trusted private network. Local history is a bounded source record, not a full backup or isolated worktree. Available as a signed in-app update; new users install the current desktop download, then update.

• Mistakes were made, that's on us. Corrected the native update restart handoff so the previous backend is stopped through the normal exit lifecycle before the app reopens. Install the fresh full version to receive this launcher fix.
• Reports now open directly into their own Kanban board: Open, In Progress, and Completed. Drag cards between stages, search the selected report, and open a card in a pop-out layer without losing your place.
• Dark lanes and alternating horizontal card gradients follow your selected palette. Count badges precede each lane title; stable ACT references sit beside clear task titles. Each lane scrolls independently, and status changes retain their history.
• Codex workspace-requirements failures now identify the provider error and offer retry or reconnect guidance. Saved work is retained; this does not claim to repair an upstream Codex service failure.
• Install the fresh 0.7.3 package for your platform to receive this update and the corrected launcher. Saved settings, chats, and projects stay in application data. Signed small updates remain supported for future compatible releases.

• Mistakes were made, that's on us. Fixed a Windows startup path bug that could prevent the app from opening, and corrected display polling that could crash Linux desktops. Fresh installers include the repaired native launcher.
• Codex connections recover when an installed ChatGPT or Codex app moves its bundled executable. Browser pickers show application icons, including the selected browser.
• Chats can draw on short historical snippets from other chats in the same project, with the source chat clearly named. Hidden worker chats and other projects are excluded.
• Consecutive edits to the same file collapse into one expandable entry while preserving each saved diff and restore check.
• Close or reopen reports, delete reports while keeping their Action Items, and delete Action Items with confirmation. Reports remain reachable from project settings.
• Debug an issue prepares a focused chat for reproduction, diagnosis, a saved finding and verification of the fix. This guided workflow uses your selected model and existing permissions.
• Fresh 0.7.2 installers and compatible signed small updates are available. If an older Windows or Linux installation cannot open reliably, install the new full version.

• Improved resilience during bursts of chat updates and fixed a Unicode handling bug that could leave a response stuck. Recoverable request failures stay isolated so other chats can continue. A bounded private log helps diagnose service issues without recording prompts or credentials.
• Multi-Model Mode instructs the Primary to delegate implementation to the selected Secondary and review its work. The compact pill keeps its styling while busy; hover or keyboard focus reveals model details with two live token charts side by side: the current task, including review continuations, and the entire chat, with Primary and Secondary role totals across earlier tasks and models. Each Secondary assignment counts once; context-compaction requests count toward lifetime usage rather than either chart. Model behavior can vary.
• An optional encrypted OpenRouter Management Key shows your account credit balance. Without a Management Key, an ordinary API key reporting no budget (not a known account balance) shows ∞ with the exact help: “No budget allowance set, usage is based on Credit Balance, configure the optional Management Key to see your current Credit Usage.” Use the browser picker to open Management Keys; click a known balance and choose a browser to open credits.
• At low allowance, Codex additional credits appear as a nominal USD equivalent; positive allowance remainders below 1% show <1%. Claude Code extra usage is labeled as spending and a monthly cap, not a prepaid balance—the live usage feed does not provide prepaid balance.
• Account reads use no model tokens. Regular model work can still consume provider allowances or credits.
• Existing 0.7.0 installers remain the baseline. This is a small signed native update; no new full installers are included.

• GalaxyDesk has a new native foundation: a Tauri desktop host and Rust agent engine replace Electron and the bundled Node service. Your workspace keeps its personality while the machinery underneath changes completely.
• Multi-Model Mode pairs a Primary model with a Secondary worker. Let one guide and review while the other handles delegated work, with clear roles and separate usage reporting.
• Native services handle project files, commands, providers, MCP, cloud storage, private connections, browser control and local OCR. Managed work has explicit cancellation and shutdown paths.
• Large saved file changes move into local archives, keeping their contents available without sending the whole archive to the chat interface at startup.
• Signed small updates verify the complete target before activation and retain recovery if a new version fails to start. This core replacement uses a new full installation; subsequent compatible updates can contain only changed content.
• Your colour palette now reaches access explanations and chat details. Multi-Model controls float above the composer, leaving the writing area clear.
• Fresh native packages cover Apple Silicon and Intel Macs, Windows x64 and ARM64, and Linux x64 and ARM64. Existing saved projects, chats and settings carry forward.

• Change an action item’s status directly from its row in the project worklist or a saved report. Choose Open, In progress, Blocked, Completed or Dismissed.
• Manual completion needs no model response or mandatory note. Changes are recorded as user updates; items can be reopened, and assistant-reported fixes still require verification evidence.
• Windows Codex sessions now use a native-compatible read policy instead of the restricted-read allowlist rejected by the unelevated sandbox. Project write limits and GalaxyDesk tool approvals remain in place.

• Removed GalaxyDesk’s fixed 15-minute cutoff for Codex tasks. Longer work can continue until completion, cancellation, or a real connection error.
• Queued prompts now capture their own attachments and clear the input tray immediately. Files added for a later prompt are preserved, and failed submissions restore the original attachments.
• Queued attachments show their count and filenames on hover, making it easier to see what belongs to each prompt.

• New steering messages appear at the point they are sent, with subsequent assistant replies and tool activity beneath them. The order stays correct when the conversation is reopened.
• Multiple steers keep separate positions in the conversation. A rejected steer is removed from the timeline and its queued prompt remains available to retry.

Older conversations without steering-position markers keep their saved layout.

• GalaxyDesk showed Full access while telling Codex its session was read-only. That contradiction caused Codex to refuse work you had already authorized. This was our integration mistake, not your settings.
• We corrected the permission mapping so Full access permits project writes. Read-only mode remains restricted, and Ask mode still routes approvals through GalaxyDesk.
• You should not have had to keep asking the assistant to try again. We're sorry for the interruption and frustration.

• Malformed tool arguments now return a useful error to the model instead of ending the task. The model can correct the call and continue in the same conversation.
• Invalid arguments are rejected before the tool runs. Completed work stays intact, and repeated failures still use the existing loop-review safeguards.

• Penelope, Merlin, Clippy and Jeeves join Violet as pre-installed companions. Choose yours in Configure → Pets.
• The pet library uses three cards per row. A live 1×–2× size slider changes your companion without enlarging chat bubbles.
• Verified context and output limits for Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna, with documented API effort choices. Installed coding engines retain their reported effort options.
• A new Pets page introduces all five companions. The comparison chart reflects current competitor documentation, including pet customization, cloud sessions and parallel agents.
• Fresh 0.6.0 installers for Windows, macOS and Linux include all recent additions. Existing users receive a smaller signed update.

• The old 30-round limit on an entire task is removed. The default is now 30 app-managed tool calls between visible assistant updates, resetting whenever new reply text appears.
• Execution settings let you raise the limit for unusually long stretches of tool work. Thinking text does not reset it.
• Loop protection checks repeated results, short repeating cycles and repeated failures. It shows the evidence and asks whether to stop or continue; ordinary running-job polls are excluded.
• Appearance now has independent UI/control and chat-text size sliders, with immediate previews and saved preferences.

• ChatGPT sign-in now offers an Open with picker for installed browsers, including Brave, plus an explicit system-default choice and Copy sign-in link.
• GalaxyDesk remembers your last browser selection without changing the operating system’s default.
• Temporary Claude 500/529-style failures receive up to two delayed retries when the failed attempt produced no reply or tool activity.
• Provider errors retain the selected model name and show a clear explanation instead of an internal synthetic label or a contradictory success message.

A persistent provider outage can still prevent completion. Work is not automatically replayed after tools have started.

• The waiting banner now counts down live without refreshing the chat or interrupting your draft.
• Smart auto-complete recognizes Claude’s fully refreshed allowance when the next reset timestamp is not yet reported, then resumes the existing task after the account and task checks pass.
• Allowance checks now run once a minute. Check now immediately refreshes the selected chat’s provider.
• The default pause cutoff is now 5%, replacing the previous 2% default. Other saved thresholds are preserved.

• Cloud Storage settings now follow the Git panel: provider logos, clear connection states, inline configuration, and one Save & test action.
• Claude Code questions now open a dedicated answer form with single-choice, multiple-choice and written answers. Your actual answers return to the running session.
• Pending questions survive a page reload. Skip and Stop cancel them without inventing an answer.

• Pet bubbles now name their project and chat, update with the latest reply, and stack for concurrent conversations. Open or dismiss each chat independently.
• Connect Dropbox, Google Drive, Box, OneDrive, FTP/FTPS or SFTP with your own credentials, stored in the OS-encrypted vault.
• Browse cloud folders, search the current folder, follow breadcrumbs, and select multiple files. Reference files in a chosen chat, edit local working copies, publish explicitly, or confirm deletion.
• Upload from your computer through the sidebar cloud browser. The chat attachment button now offers Cloud and Local when connections are configured.
• A searchable emoji picker adds personality at your cursor, with recent choices and no automatic send.

This release uses user-supplied access tokens or server credentials. Reference supports UTF-8 text up to 2 MB; editable copies and uploads support files up to 25 MB. Native Google documents can be referenced through supported text exports, but must be exported before round-trip editing.

• Replies and tool activity now appear in their original sequence, including separate Claude Code response blocks.
• Browser screenshots appear where they were captured, as compact thumbnails. Consecutive captures form a grid; open one to view it larger and move between images.
• New conversation ordering is saved when you reopen the chat. Older screenshots also use thumbnails.
• The comparison page now offers a searchable 30-feature matrix across GalaxyDesk, Codex, Claude Code, Cursor, Cline, Aider and Devin, with sources and a filter for GalaxyDesk gaps.

• See remaining Codex and Claude Code account allowances in the bottom bar, with separate five-hour, weekly or other reported windows and reset-time tooltips.
• Usage checks run every five minutes through the official local clients without sending a model prompt or using inference tokens.
• Palette choices now carry through sidebar icons, skill indicators, project selections, chat links, inline code, tool activity, sources and diff controls, with smooth colour transitions.
• Queue follow-up prompts across providers, edit them, and drag to reorder. Five queued rows stay visible before the list scrolls. Steer a live Codex turn, or interrupt and redirect Claude Code without losing the saved conversation.
• Optional Smart auto-complete pauses integrated chats near your chosen remaining percentage, then confirms the allowance has reset before continuing the original task.
• Automatic continuation preserves saved progress and appears simply as “Resuming…” in the conversation. Cancel it at any time; changing the task’s model, folder or access also cancels it.

Enable Smart auto-complete in Configuration → Usage limits. The app must be open and unlocked to continue. Readings depend on the installed CLI and signed-in account; periodic checks can lag fast consumption. Continued work uses your normal model allowance.

• Full access now honours your choice for Claude web search and page fetching, shared-browser actions, and connected MCP tools.
• Ask mode, read-only restrictions, and stricter chat-level permissions still apply.
• A chat must inherit the project’s Full access or select it explicitly; an existing Ask override is kept.

• Long Claude Code tasks no longer hit GalaxyDesk’s fixed 15-minute turn limit.
• Tool activity and time spent waiting for approval can continue until the task finishes or you stop it.
• Stop, app locking, and connection-error handling remain in place.

• Refined interface controls, keeping the GalaxyDesk workspace and starfield you know.
• Keep the original GalaxyDesk palette or choose Strawberry Orange Burst or Lime. See colours crossfade as you choose; save to keep the palette.
• Open files in an installed editor with app icons, reveal them in their folder, or choose a different application through the system picker.
• Text attachments open as working copies, keeping the original chat attachment intact.

• Magnitude joins the local model providers, with model discovery and reported context, tool, vision, and reasoning-effort capabilities.
• Git is off by default. Connect GitHub, GitLab, Bitbucket, or Azure DevOps, then enable it. Tokens use protected storage and stay out of model prompts.
• Clone repositories into projects, inspect diffs and history, initialize repositories, manage branches, stage specific files, commit, fetch, pull, and push through a linked provider. Model-requested pushes ask for approval.
• Configuration now has a left-hand menu with icons, a fixed navigation area, and independently scrolling settings.

• Fresh 0.4.0 downloads for Windows, macOS, and Linux include every improvement so far. Existing users can update in the app.
• Violet comes pre-installed. Enable Pet Assistant and she is ready to keep you company—no files to find or packages to import.
• Your selected pet and on/off preference carry forward. Violet is the default only when you have not chosen a companion.
• Violet stays available as an included pet. You can still install and switch to your own companions.

• Optional automatic update checks every 15 minutes, with a blue UPDATE button to download and restart in one click.
• Pet Assistant: install Codex-format pet packages, choose an installed pet, and drag your companion around the desktop.
• Response bubbles show the beginning of a reply, mark completion with a checkmark, and open the matching conversation when clicked.
• Pet Assistant hides and clears its bubble when the workspace is locked. Pet choices and positions are saved across launches.
• Full access chats can open requested localhost previews directly. Other access levels receive an inline approval prompt; other private origins remain restricted.

• Improved tool recovery when continuing older or compacted conversations with local models.
• Rejected calls receive precise corrections, and historical examples are kept distinct from executed actions. Invalid calls remain blocked.

• Reliable tool execution for Qwen 2.5 Coder, with structured recovery for supported local servers when a model prints a call instead of making one.
• Full Ollama model context by default, with an explicit custom limit when you want one. Correct Gemma 4 12B capacity: 256K.
• Refreshed model capacities, accurate context after switching providers, and stable Claude Code compaction budgets.
• Original user requirements retained alongside summaries, clearer activity counts, and compatibility with current Codex App Server permissions.

• Action Items appear under their owning project only when present, with clear Project reporting and Project workflow skill labels.
• Persistent project reviews and Action Items, including follow-up reviews across models.
• Encrypted project connections and an optional password/startup app lock.
• Quick setup for installed tools, plus Claude Code in regular chats and optional native terminal sessions.
• Six platform builds with signed, small shared-code updates for compatible future releases.

• A browser view shared with the assistant, plus configurable MCP servers.
• Precise edits, reviewable diffs, guarded file restoration, and commands that keep running.
• Optional subagents with concurrency controls, and an Outputs panel for current work.
• Lifetime, project, and response token accounting, with reported usage and estimates kept separate.

• Global and project-specific memory that you can add, edit, and remove.
• Adjustable context layers, with more room protected for the current conversation and tools.
• A bundled model-capacity catalog and upgrade-safe settings.

• Local and cloud providers, assistant and user profiles, Markdown bios, and reusable skills.
• Project chats, access controls, coding tools, command activity, and collapsed file diffs.
• Automatic and manual compaction, original-chat search, dropped attachments, and saved OCR/image records.