v0.7.1

Changed

• 🤖 More reliable Codex agent connection. The Codex agent now handles large and unexpected output more gracefully, with better error messages when the connection drops or something goes wrong. You'll see a clear explanation instead of a silent failure.
• 🔄 Live command output for Codex agent. When the Codex agent runs commands or edits files, you can now see the output streaming in real time instead of only after the step finishes. Long output is automatically trimmed to keep things readable.
• 🤖 Codex agent handles approval requests. The agent now responds to approval prompts from newer Codex versions automatically, based on the permission mode you've configured.

v0.7.0

Added

• 🤖 Coding agents. You can now connect coding agents as AI backends. Supported agents include Codex, Claude Code, Cursor, Grok, and OpenCode. Configure agent profiles from the new Agents tab in admin settings, and they show up as selectable models in the chat. Each agent runs in your workspace with full tool access, and conversations can be resumed where you left off.
• ⚙️ Agents admin panel. A new Agents section in Settings lets you add, edit, and remove agent profiles. You can configure the command path, model list, approval and permission modes, and see at a glance whether each agent is installed and ready. Models can be left empty to auto-detect what the agent supports.
• 🎙️ Voice dictation in the terminal. The mobile shortcut bar now has a microphone button that lets you dictate text directly into the terminal using speech recognition.
• ℹ️ System info modal. A new "System info" option in the sidebar menu shows your machine's hostname, OS, CPU, memory, disk, and running processes in a dedicated modal instead of on the welcome screen.

Changed

• 🏠 Redesigned welcome screen. The home page now highlights a "Continue" section that picks up your most recent workspace with context about what was happening (active tasks, running processes). Recent workspaces also show status signals so you can see which ones have work in progress.
• 🏷️ Renamed to Computer. All visible references to "cptr" in the UI, notifications, and documentation now read "Computer" for clarity.
• 🧹 Code formatting cleanup. Whitespace and formatting improvements across frontend components.

v0.6.2

Added

• 🗄️ Memory vault. Memories are now stored as organized Markdown files instead of a flat list. Related memories can be grouped under headings, linked together across files, and browsed from the new memory file browser and search endpoints. The AI can also link, move, split, and merge memories to keep things tidy over time.
• 🔍 Memory search and file browsing. New API endpoints let you search across all stored memories, list memory files with their headings, read individual files, and run a safety review of the vault.
• 🧠 Smarter memory recall. When building context for a conversation, the AI now pulls in relevant memories based on what you are actually talking about, not just the full baseline list. Mentioned files and recent messages help guide which memories surface.
• ⚙️ Per-model context compaction threshold. You can now set a custom compaction threshold for each model from the Models admin page. Useful for models with smaller context windows that need to compact earlier. The global threshold still acts as an upper limit.

Changed

• 💬 Queued messages follow the active branch. Queued messages now only appear if they belong to the branch you are currently viewing, instead of showing every queued message regardless of branch.
• 📝 Conversation summaries saved in the right place. When a long conversation gets compacted, the summary is now stored on the user's message instead of the assistant's, so it sticks around even if you regenerate a response.
• 🔄 Pending input handling improved. When processing queued user messages, the system now cancels any unfinished parent response before starting the new one, and picks the right model more reliably.
• 🧹 Workspace navigation edge case fixed. Switching between workspaces with deep links no longer occasionally processes navigation actions before the workspace has fully loaded.

v0.6.1

Added

• 🔍 Firecrawl web search. Added Firecrawl as a web search provider. Configure your API key from Settings > Web, or set FIRECRAWL_API_KEY as an environment variable. Firecrawl slots in between Brave and DuckDuckGo in the automatic provider order, and also works with a custom self-hosted endpoint.

Changed

• 📂 Workspace paths are normalized. Paths like ~/Projects/myapp and /Users/you/Projects/myapp now resolve to the same workspace instead of creating duplicates. Existing duplicates are cleaned up automatically when you open or save a workspace.
• 🔔 Clicking a browser notification opens the chat. Desktop notifications for completed tasks now take you straight to the conversation when you click them, instead of just bringing the window to the front.
• 🏷️ Workspace names display consistently. Folder names shown in the sidebar, search results, automation list, and notifications now all use the same logic, so you see the same label everywhere.
• 🧹 Pending messages follow the right branch. When queued messages or background sub-agent results come in, they now attach to the correct point in the conversation instead of occasionally landing on the wrong branch.

v0.6.0

Added

• 🧠 Memory. The AI can now remember things about you and your projects. Memories are stored per user and per workspace, and are automatically included in future conversations. You can view, edit, and delete memories from the new Memory tab in Settings. Admins can also turn on background memory review, which lets the AI quietly pick up on preferences and patterns as you chat.
• 🎨 Image generation and editing. Ask the AI to create or edit images right from the chat. Generated images are saved to your workspace and displayed inline in the conversation. Supports any OpenAI-compatible image API. Configure from the new Images tab in admin settings.
• 🔀 Background sub-agents. Sub-agents can now run in the background. The AI kicks off a task, keeps chatting with you, and brings the results back when the background work is done. Great for long-running research or multi-step tasks you don't want to wait on.
• 🔍 Better chat search. Searching your chats now looks at chat IDs, titles, summaries, and message content all at once, with smarter ranking. Exact and prefix matches on titles and IDs show up first, then summary matches, then message content. You can also filter by workspace and choose whether to include sub-agent chats.

Changed

• 🖼️ Generated images show inline. Images created by the AI now appear as proper image previews in the chat instead of raw file paths or JSON. The tool call output also shows a cleaner summary instead of a wall of data.
• 🏷️ Background sub-agents are labeled. When the AI runs a task in the background, the tool call label now says "Background sub-agent" so you can tell it apart from a regular sub-agent at a glance.
• 🧹 Code cleanup. Formatting and style improvements across the codebase for better readability.

v0.5.6

Changed

• 🧠 Smarter system prompt. The AI now knows more about the machine it's running on, including the hostname, architecture, shell, and whether it's inside a container. This helps it give more relevant answers when you ask about your system or run commands.
• 📝 More template variables for custom prompts. If you customize your system prompt from the Models admin page, there are new placeholders you can use: runtime environment, hostname, platform, architecture, shell, home directory, and cptr version.

v0.5.5

Fixed

• 🔌 Socket events no longer missed on slow connections. Components that listen for real-time updates (sidebar chat list, chat panel, notifications) previously relied on polling to check if the socket was ready. If the socket connected slowly or reconnected, events could be silently dropped. Listeners are now registered centrally and automatically re-attached whenever the connection comes back.
• 🔄 WebSocket connection is more resilient. The socket now prefers a direct WebSocket connection and falls back to polling only when needed, reducing latency and improving reliability on unstable networks.

Changed

• 📂 Git status refreshes are debounced. Rapid filesystem changes (like saving multiple files at once) no longer trigger a flood of git status checks. Updates are batched with a short delay so the UI stays responsive.

v0.5.4

Added

• 📱 Full PWA support. cptr can now be installed as a standalone app on phones, tablets, and desktops. Includes offline caching, an offline fallback page, home screen shortcuts (New Chat, Open Workspace, New Note, New Terminal, Search), and a service worker that keeps static assets available when the server is unreachable.
• 📤 Share target. Share files, text, or links from other apps directly into cptr. On mobile, use the system share sheet to send content straight to a chat.
• 📂 File handling. Opening supported file types (code, documents, images) with cptr now imports them into your workspace with a folder picker.
• 📊 Context usage indicator. The chat panel now shows how full the context window is, so you can see at a glance how much room is left before compaction kicks in.
• 🗂️ Workspace picker for imports. When importing shared files or opening files from outside the app, a workspace picker lets you choose where to save them.
• 🔧 PWA settings tab. A new tab in Settings shows your install status, lets you check for service worker updates, and clear the offline cache.
• 🖥️ Status modal. A new status indicator in the chat panel shows server connection state and context usage at a glance.

Changed

• 🌐 Perplexity base URL is now configurable. You can point the Perplexity search provider at a custom endpoint (like a LiteLLM proxy) from Settings or via the PERPLEXITY_BASE_URL environment variable.
• 🤖 Simplified default system prompt. Removed the instruction that told the AI to always create a plan before acting. The AI now helps directly unless you ask it to plan.
• 📂 File uploads no longer overwrite existing files. Uploading a file with the same name as an existing one now automatically adds a number suffix instead of replacing it.
• 🌍 New translation keys. Added PWA, share, file handling, and status labels across supported languages.

v0.5.3

Added

• 🔊 Auto-stream TTS. New toggle in Audio settings that automatically reads AI responses aloud as they stream in, without needing to manually enable playback each time.
• 🔀 Branch management. You can now rename and delete branches directly from the Git panel. Right-click (or tap the action button) on any branch to see your options.
• 🔍 Branch search. The branch picker now has a search bar so you can quickly find branches in large repos.
• 💾 Stash awareness when switching branches. If you have uncommitted changes and try to switch or create a branch, you'll be asked whether to bring the changes along or leave them behind.

Fixed

• 🧩 MCP tool server errors are now helpful. If the MCP package isn't installed, you'll get a clear message telling you how to install it instead of a confusing traceback.
• 🔧 AI tool calls no longer break on empty arrays. Fixed a subtle issue where providers returning empty tool call lists could cause errors during streaming.
• 🗂️ Git file list no longer shows duplicates. Files that were both staged and modified used to appear twice in the changed files list. They now show as a single entry with the correct status.

Changed

• 📋 Dropdown menus support action buttons. Menu items can now have a secondary action icon on the right side, used for things like branch context menus.
• 📂 Home page lists are trimmed. The recent files and folder suggestions on the welcome screen now show at most 5 items each to keep things tidy.
• 📦 Added an "all" install extra. You can now pip install cptr[all] to get every optional dependency (MCP, document support, PAM) in one go.
• 🔊 Sticky save button in Audio settings. The save button in Audio settings now stays visible at the bottom of the panel as you scroll.
• 💾 Stash includes untracked files. When stashing changes, new files that haven't been committed yet are now included automatically.

v0.5.2

Fixed

• 📋 Artifact cards no longer stretch or clip text. The small preview cards that appear when the AI creates a file now have a consistent, fixed height. Long titles and previews are properly truncated instead of pushing the card taller or overflowing.