What's new in Pandore.

Your daily summary now takes action, and Pandore updates itself on Windows.

• From the daily summary, hand a task to an agent or prepare a reply to an email in one click. Nothing goes out until you've read it.
• On Windows, new versions now arrive on their own, and Pandore is better at recognising a Claude Code you've already installed.
• A failed session says plainly what happened, and an interrupted one can be resumed in one click. Pandore no longer re-runs a brief when nothing has changed in the folder.
• Once an iPhone is linked, Pandore opens when you log in, without a window, so your phone can always reach it. You can turn this off in the iPhone settings.
• The interface is now translated into English.
• To know how many installations are in use, Pandore sends its version and the type of computer once a day, with no identifier. You can turn this off in the settings.

A faster, clearer interface, notifications when your agents need you, and your teams called with @.

• A window closer to your system, fewer frames, and screens that open at once. Light and dark mode now change everywhere in one go.
• ⌘K (Ctrl+K on Windows) to search, open or ask your agents a question, and ⌘1 to ⌘9 to move between sections.
• A notification when an agent finishes a long task, asks a question or waits for your approval.
• Type @ and a team's name: all its members answer. The Agent teams tab leaves projects: the chat is enough.
• The daily summary finally shows all its sections: to do, unread emails, watch out.
• Tasks check off instantly, and a deletion can be undone in one click.

Send your emails from the conversation, form an agent team by calling them with @, and see your unread mail in the daily summary.

• An agent prepares an email: it appears in the conversation, you review it, edit it, and it leaves when you click "Send". An agent never sends anything on its own.
• Call several agents with @ in one message: they form a team, answer in turn, and each one reads the previous answer. The project Agents tab is gone: the chat is enough.
• The daily summary now also covers your latest unread emails, without the advertising.
• Local models can now work on project files, and read your mail, calendar and contacts.
• Microsoft account: Outlook, calendar, OneDrive and contacts connect in one click on Mac.
• Right-click a project to delete it, an indicator during long tasks, and a button to sign an agent back in.

Choose what Pandore shows, and connect your Google account in one click.

• Settings › Profile: "Self-employed" or "Employee". As an employee, Pandore focuses on your projects and calendar, without clients or invoices.
• One switch per module: clients, pipeline, finances, missions, calendar, AI usage. Nothing is deleted: what you hide comes back untouched.
• Google account: click "Sign in", pick your account, and your agents read your mail, calendar, files and contacts. They prepare drafts that you send yourself.
• A different Google account for each client: its projects use it on their own.
• Your days, sales and time off are simpler to enter in Finances.
• AI settings fit on a single screen, and the daily summary is clearer.

Attach files in the chat: the agent reads them, and they are saved in the project folder.

• Add files with the paperclip, by dragging them from the Finder onto the chat, or by pasting a screenshot.
• On send, they are copied to an "Attachments" (Pièces jointes) folder in the project, without ever replacing an existing file.
• The agent opens attached files itself, PDFs and images included, and remembers them later in the conversation.
• Local models read the text of files and PDFs; a message with an image goes to your remote agent in automatic mode.
• Files stay visible under your message: one click shows them in the Finder.
• The small labels in the AI usage table are easier to read.

AI usage tracking goes visual: you can see at a glance which models are working for you, and what it costs.

• Charts show the split by model, between local and remote models, and between agents, Missions and chat.
• Each model appears with its logo: Claude, OpenAI, Gemini, and, for local models, Qwen, Gemma, Mistral, Llama and the others.
• A day-by-day chart separates what ran on your computer from what went to a provider.
• A table breaks down each model: runs, tokens used, reported cost and share of the total.
• A button lets you count everything in tokens, runs or cost.
• Old replies saved without any information no longer skew the totals.

Automatic AI setup

• Pandore automatically prepares the components Codex and Gemini need.
• The installation stays within Pandore's own data and does not change the Mac's system configuration.
• Technical downloads are verified before use.
• Commands, system paths, error codes and raw logs are no longer shown.
• The interface now uses simple steps: preparing, installing and ready.
• The "not installed" and "adapter not verified" labels have been removed from the process.

The real identities of AI platforms

• The OpenAI, Anthropic and Google logos are now clearly visible in Settings.
• Local models also show the identities of Qwen, Mistral, Meta and the other available creators.
• Old provider names are recognized automatically, to avoid abbreviations like "OA".
• Logos are bigger and easier to read in the list of AI connections.

Simpler AI settings

• Choosing the AI engine now comes down to three clear modes: Automatic, Remote models and Local models.
• Automatic mode explains how Pandore splits requests between the Mac and the cloud.
• The pros and limits of each mode are shown before you select it.
• The active setup, the chosen remote service and the number of installed local models are shown in one place.
• The local model catalog is now grouped in Settings and only opens when you need it.
• The Agents section stays focused on agents, teams and automations.

The AI model in use, in real time

• A new compact indicator appears in the top bar, next to agent tracking.
• It shows right away whether Pandore is using a local or a remote model.
• The provider and model actually in use are visible during chats, Missions and agent sessions.
• Claude Code, Codex, Gemini and local models each have their own look.
• Several calls running at once are grouped under a counter, and their details remain available on hover.
• When idle, the indicator clearly confirms that no model is in use.
• One click opens the AI Usage dashboard to see detailed usage.

AI budgets by project

• The FinOps budget is now available from a compact button in the AI Usage dashboard.
• A new cockpit brings together the cost cap, token cap, alert threshold and usage for the period.
• Each project can get its own budget in dollars and in tokens.
• Pandore shows each project's progress and flags overruns without interrupting work in progress.
• A suggested split distributes 80% of the budget based on observed usage and keeps 20% in reserve for less active projects.

The AI Usage dashboard becomes a true FinOps cockpit for local and remote models.

• Usage is now split between local models and remote agents.
• The local autonomy rate shows the share of tokens processed directly on the Mac.
• Pandore estimates the API costs you avoided by applying the remote price actually seen over the period to the local volume.
• The average remote price per million tokens lets you check the calculation.
• A per-project view shows the local and remote mix, along with the estimated savings.
• Usage that cannot be attributed stays clearly flagged instead of being sorted arbitrarily.
• The OpenAI logo now displays correctly in Pandore and on the website.

The local library becomes much more complete and easier to explore.

• 43 local models are now available to download.
• The catalog covers Qwen, Gemma, Mistral, Phi, Llama, DeepSeek, SmolLM, Liquid, Granite and OLMo.
• Several sizes are available, from models under 500 MB to high-quality models for Macs with more memory.
• Filters separate general-purpose models, coding models and reasoning models.
• The "For this Mac" view automatically hides models that are too demanding for the available memory.
• Search finds a model by name, publisher, size, license or use.
• Each model card shows the download size, recommended RAM, quantization, license and where the GGUF file comes from.

Local models reply faster and stay quiet in conversations.

• Pandore preloads the selected local model in the background at startup.
• The model stays in memory while Pandore is running, so it does not have to load again for the first message.
• Switching models frees the old one before preparing the new one.
• Switching back to remote mode or turning off local AI frees the memory it was using.
• Conversations now use the persistent local server instead of the interactive terminal.
• Startup logs, the llama.cpp logo, technical commands and the system prompt can no longer appear in a reply.

Pandore can now use a remote engine, a local engine, or both together.

• The AI settings offer three clear modes: Remote, Local and Both.
• In hybrid mode, Pandore picks the engine before each request based on privacy, the tools needed, complexity, speed and cost.
• Simple requests and sensitive content are preferably handled locally when a model is installed.
• Missions, web search, connectors, file actions and structured outputs stay with the compatible remote engine.
• The provider and model stay fixed for the whole reply, to keep the conversation and usage tracking consistent.
• Deleting the last local model automatically switches Pandore back to the remote engine.

AI engines become easier to understand and control.

• The engine and model currently in use are shown directly in Settings, under AI Agents.
• Remote and local models are clearly separated, with Claude Code, Codex and Gemini on the remote side.
• The local library now offers seven GGUF models, with their publisher's logo, uses, size and license.
• Auto mode picks, for each request, the remote or local model best suited to code, reasoning, writing or quick answers.
• The local engine is set up automatically when you download your first model, with no separate install.
• An agent, or all AI, can be turned off without deleting projects or conversations.
• A saved API key can be forgotten from within Pandore, and each local model can be deleted from the Mac.

Pandore can now run AI models directly on your Mac.

• The engine choice clearly separates remote models (Claude Code, Codex and Gemini) from local models.
• A new Local Models library offers selected GGUF models for conversation, light tasks and code.
• Pandore detects your Mac's memory, tells you whether a model is a good fit, and shows its size, uses and license before you download it.
• Downloads show their progress, and installed models can be selected for new chats or deleted.
• Local conversations run through llama.cpp and keep prompts and documents on the Mac.
• Projects now have an AI policy that will let you require local, prefer local, or explicitly allow the cloud.

Pandore gives you more control over agents, costs, memory and conversations.

• Mission Control now shows each agent's recent activity, step and status, with a control to pause the whole team.
• Each completed Mission produces a consolidated summary that brings together the result, deliverables, decisions, quality check and next action.
• AI Usage accepts token or cost caps and a configurable alert threshold for each space or client.
• Pandore Memory lets you correct, pin or forget a piece of knowledge without stopping it from working automatically.
• A chat can now branch off from any finished message to explore another direction without changing the original thread.

The chat now accepts several requests while an agent is working.

• Each new message joins a queue shown at the bottom of the conversation.
• Messages are sent automatically, one at a time, as soon as the current reply is finished.
• The queue stays with the conversation when you switch screens or restart Pandore.
• You can choose the next message, move it up or down, edit it or remove it before it is sent.
• Agent mentions stay in the prompt and call the right agents when its turn comes.
• A prompt you already sent can be put back in the message box to reword it or change its direction.

Pandore Missions now coordinates agents in parallel when their work is independent.

• Up to three agents can work on a Mission at the same time.
• The plan's dependencies are still respected: checks and delivery wait until the work they need is finished.
• The control panel shows each active agent and the task it is working on.
• AI costs and calls are still counted correctly when several agents finish at the same time.
• In projects, tickets can once again be moved between Kanban columns by drag and drop.
• Moving a ticket to another column shows up immediately and correctly updates its completion date.

Pandore now lets you know automatically when a new version is ready to install.

• Pandore quietly checks for updates at startup, then regularly while the app stays open.
• It also checks again when Pandore comes back to the foreground or when the internet connection comes back.
• The notification appears in the bottom-right corner, without you having to open Settings.
• A single click on "Install and restart" downloads the version, installs it and reopens Pandore.
• If an agent is still working, Pandore lets you choose to wait before restarting.

Pandore Memory now learns on its own, quietly, while you work.

• Conversations, Missions and agent sessions automatically build up each project's memory.
• There is nothing for you to add, confirm or set up.
• The Memory tab simply lets you see what Pandore already knows about the project and where each piece of information comes from.
• The technical elements used for memory stay hidden, including while a reply is being generated.
• Knowledge already suggested in previous versions is automatically added to the project's memory.

Pandore Memory now learns as you work, while letting you decide what becomes lasting knowledge.

• Conversations can suggest saving a decision, a constraint, a preference or a lesson learned to the project's memory.
• Agents report their lasting findings along with the session they come from.
• A new "Analyze project" button reads through the folder's files and prepares a selection of knowledge for you to approve.
• Suggestions stay separate from the active memory until you confirm them.
• Technical collection blocks are hidden from agent reports.

Pandore Memory now gives every project a lasting memory.

• Keep important goals, decisions, constraints, preferences and lessons learned in the Memory tab.
• Approve or dismiss the knowledge suggested at the end of Missions: nothing is added to the AI's context without your approval.
• Confirmed knowledge is automatically shared with the project's chat, agents, agent teams and Missions.
• Pin the key items so they keep priority when context is limited.
• When a new version is available, Pandore now shows a notification in the bottom-right corner to download it, install it and restart right away.

The AI Usage dashboard is now available in the Professional and Personal spaces.

• Track the input, output and cache tokens used by your agents.
• Analyze usage by project, client, agent and model.
• Find the most expensive sessions to understand where your AI budget goes.
• The detailed history starts with the runs measured by this version; older sessions stay visible, marked as not measured.

Give Pandore a goal with Missions, and call your agents right in the chat with @.

• Missions: describe what you want to achieve in one sentence. Pandore proposes a plan, carries it out step by step, has it reviewed, then delivers the result to you. Nothing important happens without your approval.
• Two examples to get started: turn an app idea into a real project, or prepare a priced sales proposal using your rates.
• In the chat, type @ to call one or more agents: each one replies in its own bubble, at the same time as the others.
• Agent tracking now reads project by project: what is running, what is waiting for you, what is done.
• The English interface is more complete.
• Missions work with Claude Code.

Pandore is now signed and approved by Apple: macOS no longer flags it as an unknown app.

• When you open this version, macOS may ask one last time for access to your calendars and folders. Accept: it will not ask again with future updates.

Files mentioned by agents open in one click, notifications work again, and Pandore now updates itself.

• Chat: mentioned files open in one click, and a panel shows the ones the agent changed.
• Reliable macOS notifications, with direct access to the settings when you need to allow them.
• Claude installs and connects with two buttons, no terminal needed.
• Settings save on their own, and the daily summary stays up to date.
• Connectors: email fixed, Telegram, Airtable, Asana, and calendars through an ICS link. The others are marked "Coming soon".
• Automatic updates: Pandore offers to restart when a new version is ready.