Feature map

Live conversation for the moment. A task agent for the work that follows.

Live voice conversation: Speak through your microphone and hear spoken responses through your selected output device. Gemini Live handles the voice session.

Two cooperating brains: Everyday function calls go to the desktop service. Research, writing and complex coding can be delegated to the high-level agent.

Interruption-aware audio: The audio path handles interruption and clears pending playback so the conversation can change direction. Actual responsiveness depends on your devices and connection.

Audio-file tools: The task agent can transcribe recordings, synthesize speech and play audio through configured providers. These are separate from the live microphone session.

Recognize the project, remember its other name, and keep the next step close.

Current desktop context: Capture supported foreground-window, Explorer selection, browser and editor metadata. A snapshot anchors what you meant by ‘this’.

Projects, aliases & checkouts: Find learned projects by name, preserve explicit aliases and distinguish a repository from its separate local checkouts.

Project notes & preferences: Keep a project checkpoint or an explicit preference in the appropriate local memory system, then inspect, correct or remove it.

History retrieval: Search stored task-agent conversations and recent desktop observations. A fresh voice launch does not automatically resume your old conversation.

Memory inspection & controls: Inspect entities, notes and their sources. Disable context collection separately from forgetting stored records.

Organize your real tabs, carry useful information across apps, and return to a saved setup.

Tab grouping & recovery: Group observed tabs by purpose, preserve pinned tabs and restore supported previous layouts when newer manual edits do not conflict.

Selections & page tables: Read supported selected text or HTML tables and transfer them into a file, clipboard or connected editor.

Named workspace snapshots: Save and restore supported browser/editor references and native window positions under a name you choose.

Browser inspection & actions: The task agent also has a separately configured browser tool for navigation and page actions. That runtime is distinct from the real-profile desktop adapter.

Find the file, read the screenshot, prepare a batch, and reuse the recipe.

File search & exact operations: Search within a specified root, read files, open known paths and perform supported copy, move or rename operations.

Local indexing & OCR: Index supported text and opt selected folders into local image OCR. Search results report freshness and coverage.

Image batch preparation: Convert and resize selected images into verified outputs while retaining source files. Reuse a completed job's settings with new inputs.

PDF & archive preparation: Merge PDFs, extract selected PDF pages, create ZIP archives and extract supported archives into prepared outputs.

Video & audio preparation: Trim video or extract audio with installed FFmpeg/ffprobe. Preparation runs as a job whose output can be inspected.

Timers, notes, tasks and reminders that can take the relevant page along.

Persistent timers & reminders: Create, list, update, pause, resume, snooze or cancel supported timed items with explicit dates and timezones.

Reminders with context: Attach the actual browser page or editor file reference to a reminder, note or task.

Tasks, lists & local agenda: Keep personal tasks, notes, list items and dated local events; complete or reopen the relevant records.

Recurring agent work: The task agent's scheduler supports configured one-time and recurring work with its own tool, model and delivery requirements.

Explore a codebase, investigate a failure, research sources and create useful artifacts.

Repository exploration: Read files, search contents and trace related code before proposing an implementation or explanation.

Terminal & code execution: Run permitted commands and inspect output. Code mode offers a persistent Python execution workspace for multi-step tasks.

Diagnostics & debugging: Use connected VS Code diagnostics or the Python diagnostic tools backed by installed ty, pyright or mypy.

Specialist delegation: Delegate exploration, planning or general work to configured subagents and use their findings within the larger task.

Web research & writing: Search sources, extract page text, compare evidence and write a summary or local artifact through the configured tools.

Resolve a project and recipient, prepare the actual files, and inspect delivery receipts.

Project document sharing: Resolve README and license roles to verified project files, including a known remote revision when remote retrieval is allowed.

Share selected files: Use explicit paths, an observed Explorer selection or a supported active file as the source of an attachment batch.

Account & recipient resolution: Inspect authenticated accounts, choose the sender and resolve reachable recipients without treating an open messaging window as authentication.

Per-file delivery results: Prepared attachments retain content hashes. Each file has its own result, so partial and uncertain delivery can be inspected.

Use supported Windows controls, inspect visible context and work with connected applications.

Windows & window layout: Inspect supported windows and apply available layout operations using actual window identities.

Screen capture & text: Capture a screenshot or recognize visible text locally in an identified foreground window. Model-based image interpretation needs vision support.

Audio & display controls: Inspect and adjust supported volume and brightness controls, list active audio outputs and verify a chosen default output device.

Connected editor actions: Inspect VS Code context, open a known file, insert text into a verified selection and run supported commands or known tasks.

Calculations, units & text cleanup: Use deterministic arithmetic, percentages, bill splitting and supported unit conversions. Sort lines, remove duplicates or transform supplied text without asking a model to guess the calculation.

Watch a download or build, inspect long-running work, and undo supported actions.

Inspectable background jobs: Inspect accepted, running, completed, partial, failed and interrupted work through desktop activity and job results.

Download & task completion: Subscribe to real browser download events or the exit result of an identified running VS Code task.

Completion follow-ups: Attach a notification to a known job and report completion, failure or partial outcomes when the result arrives.

Guarded action recovery: Undo supported tab organization and file preparation outputs while checking for later changes. This is action recovery, not universal conversation rewind.

Choose the task model, connect MCP tools and teach reusable ways of working.

Task-model choice: Choose supported cloud providers, compatible endpoints or a separately running local model server for the high-level agent.

MCP tool connections: Configure servers, discover the tools they expose and invoke them with the credentials and access each service requires.

Skills & Python extensions: Add reusable instructions, custom tools, lifecycle hooks, prompt templates and specialist agent definitions.

Tool permissions & profiles: Configure tool access and path rules for the agent. Approval behavior depends on the active runtime and permission configuration.

Authenticated local services: The voice and task brains communicate through A2A; desktop tools use their local service. These retained services support integrations, not a bundled chat website.